A digital twin resource allocation method and architecture

Abstract

The present application relates to a kind of digital twin resource allocation method and architecture.The method part mainly includes: analysis demand, establish clear demand index;Credit of calculation model, estimate output parameter of model;Select all the model combination scheme that meets the demand as alternative scheme;From alternative scheme, the optimal scheme is selected to be implemented, and feedback mechanism is established according to the actual use of model.The present application proposes digital twin resource allocation method and architecture, through four modules of information collection, credit calculation, allocation strategy, reward and punishment mechanism, from model, user information matching, to predict the output of model under current environment, then according to storage, network, calculation resource, reasonable allocation is carried out to model, finally, the performance of model under the resource allocation scheme is verified, and reward and punishment mechanism is established to model.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of communication technology, in particular to a digital twin resource allocation method and architecture. BACKGROUND

[0002] The development of digital economy needs the support of information technology and communication technology, and the demand of industry digital transformation also promotes the evolution of traditional communication network architecture to "cloud network integration" architecture. The communication network architecture constructed by the traditional "physical equipment plus professional network management" mode is difficult to meet the digital transformation needs of various industries such as "resource on demand, flexible management and control, and safe and reliable". Moreover, with the continuous development of the informationization process of the whole society, the demand for network delay, capacity, bandwidth and other needs also presents a rapid growth trend, and the traditional network is facing the pressure of large-scale, large-capacity data exchange and processing.

[0003] Digital twin is derived from CPS system (Cyber-Physical Systems, information physical system), CPS is a multi-dimensional complex system integrating physical environment, information network and operation, which realizes adaptive control, comprehensive perception and information service supply through deep integration and cooperation of control, communication and computer technology. It is composed of multiple elements, including complex network composed of physical devices and digital components.

[0004] But how to accurately measure user demand, establish a reasonable model resource allocation mechanism. How to make the model resource allocation scheme more efficient, how to make the combination of multiple models more reasonable. How to predict the current model use result through the past model use situation. When the model prediction result meets what kind of condition, it is considered that the model or its combination scheme is suitable for the current situation? All the above are the problems about how to reasonably allocate model calculation, transmission and network resources in the field of digital twin. Similarly, this is also the focus problem that the whole industry is eager to understand. SUMMARY

[0005] In view of the above defects or improvement needs of the prior art, the present application provides a digital twin resource allocation method and architecture, which comprises information collection, credit calculation, allocation strategy, and punishment mechanism. When a user / demand has a task to perform, it first sends a task request to the resource calling layer, which builds an LFM (Looking For Model) containing the user's task demand, resource start time and end time, and forwards the LFM to each model. The model receives the LFM and determines whether it can execute the task demand. If so, the model calculates the required resources, accuracy, time, etc. of the current parameters to form an information tag and feed it back to the resource calling layer, otherwise it is ignored. The resource scheduling layer collects the information tags and calculates the comprehensive credit of the corresponding model. According to the allocation strategy, a suitable model is selected and executed. After the model is executed, the result is transmitted to the user / demand, and the user reports the calculation error, time, etc. to the resource calling layer, and the model credit is rewarded or punished according to the actual model usage information and tag information, and updated.

[0006] The embodiment of the present application adopts the following technical solutions:

[0007] In a first aspect, the present application provides a digital twin resource allocation method, comprising:

[0008] Analyzing the demand and establishing clear demand indicators;

[0009] Calculating the credit of the model and estimating the output parameters of the model;

[0010] Selecting all model combination schemes that meet the demand as alternative schemes;

[0011] Selecting the optimal scheme from the alternative schemes to implement, and establishing a feedback mechanism according to the actual use of the model.

[0012] Further, the analysis of the demand and the establishment of clear demand indicators specifically include:

[0013] Obtaining demand description, including transmission, storage, calculation, time, error rate, and credit; wherein the transmission, storage, and calculation index descriptions each include a minimum value and a maximum value; the time index description includes the start time and end time of operation; the error rate index description includes the maximum calculation error of acceptable demand parameters; and the credit index description includes the minimum credit of acceptable models.

[0014] Further, the calculation of the credit of the model and the estimation of the output parameters of the model include: evaluating the output of the model under the current environment according to the historical output parameters of the model, and constructing a credit tag of the model; wherein the construction of the credit tag of the model specifically includes:

[0015] Obtaining resource requirements of the model in terms of transmission, storage and calculation, if the model does not require any of the resources, the resource is represented by 0, if any resource in the resource vector exceeds the resources currently available to the model, it is considered that the model is not applicable to the current network environment;

[0016] Obtaining the accuracy of the model in calculating multiple indicators of the current task, if the theoretical accuracy of the model cannot meet the requirements of the task, it is considered that the model accuracy is limited in this network environment;

[0017] When the resource requirements and accuracy both meet the requirements of the current network environment, the credit label of the model at this time is calculated; wherein, the credit label of the model is related to the input and output of the model at the historical time, the output of the model in the current network environment state is predicted through the input and output relationship of the model at the historical time, the difference between the predicted output of the model and the task demand is compared, and the condition that the model can meet the task demand is estimated, that is, the credit of the model is obtained.

[0018] Further, when the credit of the model is calculated, both the credit calculation of a single model and the credit calculation of multiple model combinations are supported.

[0019] Further, the selecting all model combination schemes meeting the requirements as alternative schemes specifically includes:

[0020] According to the requirements, as long as the output result of all models meets the requirements, the corresponding model is selected as an alternative model;

[0021] Randomly sorting all alternative models to obtain an alternative model set;

[0022] Starting from the first model, each time the model is deleted from the alternative model set, it is judged whether the remaining set still meets the task calculation result requirements, if it still meets, the deletion process is continued, otherwise it is stopped;

[0023] The last set that cannot be deleted is selected as an alternative scheme, and the alternative schemes are arranged from small to large according to the sum of transmission, storage and calculation resources;

[0024] According to the demand resources and time requirements, all alternative schemes are adjusted in time resource planning.

[0025] Further, the adjusting all alternative schemes in time resource planning according to the demand resources and time requirements specifically includes:

[0026] If the integrated resources of the model are greater than the required resources in any of transmission, storage and calculation, step-by-step and time-sharing calculation is adopted, otherwise the same time calculation is adopted;

[0027] When the step-by-step and time-sharing calculation is adopted, the models are arranged in descending order of centrality on the resources required by the demand; the first model is taken as the first time, and the second model is added to the sequence from the first model, and the process is repeated; if any resource exceeds the total after adding the next item, it is further determined that if the last model in the sequence is removed and the next item is added, the resource is less than the total resource of the original sequence, then the last model is replaced by the next item to form a sequence, and the last model is placed in the next time.

[0028] Further, when the models are arranged in descending order of centrality on the resources required by the demand, if multiple resources exceed the resources required by the demand, the item that exceeds the most is selected for arrangement.

[0029] Further, the selection of the optimal scheme from the alternative schemes to implement specifically includes:

[0030] According to the allocation strategy, the resource occupied by the model in each scheme and the credit of the model are compared, the scheme with the least resource occupation, the least calculation time and the highest credit is selected as the optimal scheme to implement.

[0031] Further, the feedback mechanism established according to the actual use of the model specifically includes:

[0032] According to the resource occupied by the selected model in the actual use and the resource description provided by the model, the accuracy of the model is adjusted; if the actual resource occupied is closer to the implementation resource description, the model accuracy adjustment is smaller, and if the difference is large, the accuracy of the model this time is reduced.

[0033] In a second aspect, the application also provides a digital twin resource allocation architecture, which comprises an information collection module, a credit calculation module, an allocation strategy module and a punishment mechanism module, wherein:

[0034] The information collection module is used to construct an LFM, and the LFM comprises task demand, resource start time and end time, task indicators and indicator accuracy;

[0035] The credit calculation module is used to determine which model output satisfies the task indicators after receiving the LFM, and determine whether the combination of the models can execute the task demand; if yes, the credit calculation module retrieves the information of the satisfied models from the data pool to construct an information tag and transmits it to the allocation strategy module;

[0036] The distribution strategy module is used for selecting a combination of multiple models according to model information in the information label meeting the credit calculation, determining how to complete a task in time and in steps under a resource limited condition, and then sending the task and the model information to a probability model and a simulation model to call the model combination to complete the task calculation.

[0037] The punishment mechanism module is used for rewarding and punishing the model credit according to the comparison between the actual model use information and the model information label information.

[0038] Compared with the prior art, the beneficial effects of the present application are that:

[0039] (1) The present application proposes a digital twin resource distribution method and framework, which comprises four modules of information collection, credit calculation, distribution strategy and punishment mechanism. Starting from model and user information matching, the model output under the current environment is predicted, then the model is reasonably distributed according to storage, network and computing resources, and finally the performance of the model under the resource distribution scheme is verified to establish a reward and punishment mechanism for the model.

[0040] (2) The present application proposes a model combination strategy based on user task and current environment resource, and a model step-by-step and time-sharing operation mechanism. By matching with the user task demand, the model combination mode meeting the environment resource is selected, then through different model combinations, the step-by-step and time-sharing calculation method is adopted to make the models running at the same time not exceed the provided transmission, storage and computing resources, so as to complete the user task as quickly as possible. BRIEF DESCRIPTION OF DRAWINGS

[0041] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments of the present application. Obviously, the drawings described below are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0042] Figure 1 A digital twin resource distribution method flowchart provided for the embodiment 1 of the present application;

[0043] Figure 2 A user demand tensor and a model tensor schematic diagram provided for the embodiment 1 of the present application;

[0044] Figure 3 A credit label construction flowchart provided for the embodiment 1 of the present application;

[0045] Figure 4 A step 300 extension flowchart provided for the embodiment 1 of the present application;

[0046] Figure 5A digital twin resource allocation architecture schematic diagram provided for the embodiment 2 of the present application;

[0047] Figure 6 A model combination selection schematic diagram provided for the embodiment 2 of the present application. DETAILED DESCRIPTION

[0048] In order to make the purpose, technical scheme and advantages of the present application more clear, the present application is further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0049] The present application is a system architecture of a specific function system, so in the specific embodiments, the functional logical relationship of each structure module is mainly described, and the specific software and hardware implementation is not limited.

[0050] In addition, the technical features involved in each embodiment of the present application described below can be combined with each other as long as they do not conflict with each other. The present application will be described in detail below in combination with the drawings and embodiments.

[0051] With the process of digitization and intelligentization, the traditional communication network will realize machine automatic, intelligent management and control of optical network in the future, but the management and control scheme formed based on machine intelligence may not be reassuring, and it is not dare to directly issue instructions to the equipment side. Based on this situation, it is necessary to establish a twin optical network based on digital twin technology, simulate the operation mechanism of the optical network, real-time synchronize the operation state of the optical network, verify the management and control scheme constructed by the machine in the twin network, and accurately and accurately issue it to the entity equipment. This can avoid the risk brought by the management and control scheme formed by the machine intelligence. Based on this, the following embodiments of the present application provide a digital twin resource allocation method and architecture. The implementation of digital twin model parameter classification realizes the calling of model resources in the digital twin platform.

[0052] Embodiment 1

[0053] As shown in the figure, the present embodiment 1 proposes a digital twin resource allocation method, which comprises the following steps. Figure 1 Step 100: analyze the demand and establish clear demand indicators.

[0054] Step 200: calculate the credit of the model and estimate the output parameters of the model.

[0055] Step 300: select all model combination schemes that meet the demand as alternative schemes.

[0056] Step 300: select all model combination schemes that meet the demand as alternative schemes.

[0057] Step 400: select the optimal solution from the alternatives to implement, and establish a feedback mechanism according to the actual use of the model.

[0058] As Figure 2 shown, the user demand tensor and the model tensor of the preferred embodiment are shown in the figure. The demand description obtained in step 100 includes six aspects: transmission, storage, calculation, time, error rate, and credit (or reputation). Among them, the transmission, storage, and calculation index descriptions each include a minimum value and a maximum value; the time index description includes the start time and end time of the operation; the error rate index description includes the maximum calculation error of the acceptable demand parameters; and the credit index description includes the minimum credit of the acceptable model. Specifically, the user / demand description for the above six aspects can be represented as

[0059] CR = (Transmit_v1, Transmit_v2, Store_v1, Store_v2, Calculate_v1, Calculate_v2, T1, T2, Acc, Bri), where Transmit_v1 and Transmit_v2 represent the minimum and maximum values of the transmission index, Store_v1 and Store_v2 represent the minimum and maximum values of the storage index, Calculate_v1 and Calculate_v2 represent the minimum and maximum values of the calculation index, T1 and T2 represent the time when the model runs from T1 to T2, Acc represents the maximum calculation error of the acceptable demand parameters, and Bri represents the minimum credit of the acceptable model. For example, CR = ((Bandwidth, (100, 200)), (Memory, (1, 2)), (Gpu, (1, 2)), (2022.3.30.9.15, 2022.3.30.9.16), 0.3, 0.5) represents that the task provides a bandwidth index between 100-200bps in the transmission demand, a memory index between 1-2g in the storage aspect, and a Gpu number between 1-2 in the calculation aspect. It starts at 9:15 on March 20, 2022, and ends at 9:16 on March 20, 2022. The relative error of the demand parameter is within 0.3, and the credit of the acceptable calculation model is above 0.5.

[0060] Correspondingly, the model running index is divided into time Model i _T, error rate Model i _Acc, calculation Model i _Calculate, storage Model i _Store, transmission Model i_Transmit. i represents the i-th model combination solution. When the model running time is less than the time difference described by the user / demand, the computing, storage, and transmission resources are between the minimum and maximum values ​​of the corresponding resources, and the error rate is less than the error rate required by the user / demand, the selected i-th solution is considered to meet the user / demand requirements. The above conditions can be expressed as:

[0061]

[0062] In step 200 of this preferred embodiment, the output of the model in the current environment can be evaluated based on the output parameters of the model at historical moments to construct a credit label for the model; Figure 3 As shown in FIG, constructing the credit label of the model specifically includes the following steps.

[0063] Step 201: Obtain the resource requirements for transmission, storage, and computing required by the model. If the model does not require any of these resources, the resource is represented by 0. If any resource in the resource vector exceeds the resources currently available for the model, the model is considered unsuitable for the current network environment.

[0064] Step 202: Obtain the model's accuracy when calculating multiple indicators for the current task (indicating the model's theoretical accuracy under the current environment). If the theoretical model accuracy does not meet the task requirements, the model accuracy is considered to be limited under this network environment. If only one indicator is used, it is still represented as a vector.

[0065] Step 203: When both the resource requirements and the accuracy meet the requirements of the current network environment, the credit label of the model at this time is calculated. The credit label of the model is related to the input and output of the model at the historical moment. The output and input relationship of the model at the historical moment is used to predict the output of the model under the current network environment. By comparing the difference between the model's predicted output and the task requirements, it is estimated whether the model can meet the task requirements, that is, the credit of the model is obtained. It should be noted that the output-input relationship of the model at the historical moment reflects the changes in the model performance under different historical environments. The state difference between the inputs at different moments in the historical information reflects the changes in the network between the current moment and the previous moment. By ensuring that the difference in the network environment between historical moments remains unchanged, it is possible to clearly predict the changes in the model performance in the future when the environment changes by the same state difference as the past network environment.

[0066] In addition, this embodiment does not limit the number of models when calculating the credit of a model. It supports both the credit calculation of a single model and the credit calculation of a combination of multiple models. For the sake of description, this embodiment is collectively referred to as the credit calculation of the model.

[0067] like Figure 4 As shown, in this preferred embodiment, step 300 can be expanded into the following steps.

[0068] Step 301: According to user / demand, among all models, as long as there is one model whose output meets the required requirements, the corresponding model will be used as an alternative model.

[0069] Step 302: Randomly sort all candidate models to obtain a set of candidate models. This set must meet the task's computational requirements. If the set still fails to meet some of the task's computational requirements, feedback is sent to the task indicating that the computational requirements are temporarily unavailable.

[0070] Step 303: Starting from the first model, delete the model from the candidate model set each time, and determine whether the remaining set still meets the task calculation result requirements. If it still meets the requirements, continue to repeat the deletion process, otherwise terminate.

[0071] Step 304: The last set that cannot be deleted is used as an alternative solution, and the alternative solutions are arranged from small to large according to the total amount of transmission, storage, and computing resources.

[0072] Step 305: among the alternative plans, adjust the time resource planning of all the alternative plans according to the user / demand resources and time requirements.

[0073] Specifically, the adjustment strategy in step 305 is as follows: If the sum of the integrated resources of the model in any aspect of transmission, storage, and calculation is greater than the resources required by the demand, then step-by-step time-sharing calculation is adopted, otherwise calculation is performed at the same time. When step-by-step time-sharing calculation is adopted: the models are arranged from high to low according to the centrality in terms of resources exceeding the demand; (It should be noted that when the model has multiple resources exceeding the demand in transmission, storage, and calculation, it is first arranged according to the maximum number of resources exceeded.) Starting from the first model as the first moment, the second model is added from the first model as the moment sequence, and so on. If any resource exceeds the total after adding the next item, further judgment is made: if the last model of the sequence is removed and the next model is added, the resources are less than the total resources of the original sequence, then the last model is replaced with the next model to form a sequence, and the original last model is placed in the next moment.

[0074] In this preferred embodiment, the "selection of the optimal solution from the alternative solutions for implementation" in step 400 specifically includes: comparing the model resource usage and the model credibility of each solution according to the allocation strategy, and selecting the solution with the least resource usage, the least computation time, and the highest credibility as the optimal solution for implementation. It should be noted that the proportion selected in this step is determined by the current environment. For example, if the current environment has sufficient resources and ample computation time, the solution with the highest credibility is selected. Refer to the following formula:

[0075] V=(α1 / time)*(α2 / transmit)*(α3 / store)*(α4 / compute)*α5brief

[0076]

[0077] α1+α2+α3+α4+α5=1

[0078] Formula (1) represents that when considering the scheme value V, the calculation time, transmission, storage, computing resources and model credit of the scheme need to be considered comprehensively, wherein according to different attention proportions of the calculation time, transmission, storage, computing resources and model credit, the proportions are divided according to α1:α2:α3:α4:α5, formula (2) describes that the proportions α1-α5 depend on the proportion of the resources consumed by the current scheme and the required resources. Formula (3) describes that the sum of α1-α5 is 1, which is convenient for calculation.

[0079] In the preferred embodiment, the step 400 of "establishing a feedback mechanism according to the actual use of the model" specifically includes: comparing the resources occupied by the selected model in this actual use with the implementation model resource description (i.e. the previously calculated and estimated occupied resources), and adjusting the model accuracy; the closer the actual occupied resources and the implementation resource description, the smaller the model accuracy adjustment, and if the difference between the two is large, the accuracy of the model this time is reduced.

[0080] It should be noted that the above steps do not have a strict sequence limitation, for example, steps 200 and 300 can be performed simultaneously or in cross, and all steps in the embodiment can be used in any combination as long as they do not violate the logic.

[0081] In summary, the embodiment proposes a digital twin resource allocation method, which collects information, calculates credit, allocates strategy, and establishes reward and punishment mechanism. Starting from model and user information matching, to predicting model output under current environment, finally allocating resources to model reasonably according to storage, network and computing resources, and finally verifying the performance of the model under the resource allocation scheme, and establishing a reward and punishment mechanism for the model. In addition, the embodiment proposes a model combination strategy based on user task and current environment resources, and a model step-by-step and time-sharing operation mechanism. By matching with user task demand, selecting a model combination mode that meets the environment resources, and then through different model combinations, using a step-by-step and time-sharing calculation method, the model running at the same time does not exceed the provided transmission, storage and computing resources, and the user task is completed as quickly as possible.

[0082] Embodiment 2:

[0083] Based on the digital twin resource allocation method provided in embodiment 1, embodiment 2 further provides a digital twin resource allocation architecture, as shown in Figure 5As shown, the digital twin resource allocation architecture of the embodiment includes an information collection module (corresponding to information collection in the figure), a credit calculation module (corresponding to reputation calculation in the figure), an allocation strategy module (corresponding to allocation strategy in the figure), and a punishment mechanism module (corresponding to punishment mechanism in the figure). The above four modules are set in the digital twin platform as a resource calling layer, interact with the probability model and the simulation model, and realize the resource allocation of the digital twin. When the user / APP has a task to perform, the task request is sent to the resource calling layer, and the information collection module of the resource calling layer will build an LFM (Looking For Model, model query table) containing the user's task demand, the calculation scene, the start and end time of the resource, the task indicators and the accuracy of the indicators. After receiving the LFM, the credit calculation module first judges which model output meets the task indicators, judges whether the combination of the models can execute the task demand, and if so, the credit calculation module fetches the required resources of the current model, historical accuracy, etc. from the data pool, constructs an information label, and transmits it to the allocation strategy module. The allocation strategy module selects a combination of multiple models according to the model information in the information label constructed by the credit calculation, determines how to complete the task in time steps under the condition of resource limitation, and then sends the task and model information to the probability model and simulation model module to call the model combination to complete the task calculation. Finally, in the reward and punishment mechanism module, the model credit is rewarded and punished according to the comparison between the actual model usage information and the model information label information.

[0084] The functions of each module of the embodiment are described below. For the information collection module, for example: the task demand proposed by the user needs to be implemented, the user demand indicators (x1, x2, x3, x4, x5) are determined, the user task start time is 19:00, the calculation content and the required resources (time delay 120 minutes, bandwidth 100 bit / s, CPU 10, memory 200G, and the current user demand has no requirement for model credit) are determined, the time delay 120 minutes means that it will be completed before 21:00. Analyze the user demand to establish clear demand indicators (x1 error within 0.5, x2 error within 0.6, x3 error within 0.5, x4 error within 0.6, x5 error within 0.7), that is, CR=(0, 100, 0, 200, 0, 10, 19:00, 21:00, (0.5, 0.6, 0.5, 0.6, 0.7), 0).

[0085] The model feeds back information tags of its own attributes, for example, model one outputs x0, x1, x3, x5, calculation error 0.3, 0.2, 0.3, 0.1, time delay 20, bandwidth 80, CPU 5, memory 50; model two outputs x4, x5, x6, x7, calculation error 0.2, 0.1, 0.3, 0.2, time delay 70, bandwidth 40, CPU 5, memory 60; model three outputs x1, x2, x8, x 10 , calculation error 0.2, 0.1, 0.1, 0.2, time delay 40, bandwidth 20, CPU 4, memory 30; model four outputs x2, x3, x4, x9, calculation error 0.1, 0.3, 0.1, 0.2, time delay 30, bandwidth 30, CPU 3, memory 70; model five outputs x0, x9, calculation error 0.6, 0.7, time delay 110, bandwidth 30, CPU 6, memory 70. In the embodiment, x1-x 10 represent 10 calculated output performance indicators, for example, OSNR, BER, etc. The error represents the difference between the calculation and the true value, for example, x1 error within 0.5 represents that the difference between the calculation and the true value is less than 0.5.

[0086] Reference Figure 6 As shown in the figure, for the allocation strategy module, a model combination scheme selection strategy is performed according to a user demand indicator. For example: according to the task total demand calculation parameters x1, x2, x3, x4, x5 in example one, first, as long as one parameter in all models (models one to five) meets the output demand, it is regarded as a candidate model, according to the current task demand, model one contains x1, x3, x5, then model one can be used as a candidate model. In this way, the candidate model set contains candidate model one, candidate model two, candidate model three, candidate model four. It is judged whether the sum of the candidate models can meet the user calculation demand, if not, it is considered that the current network environment cannot support the user task demand.

[0087] Secondly, a single model is gradually deleted from the candidate set to determine the combination scheme of the model.

[0088] Model one in the candidate set is deleted, then model candidate scheme one (candidate model two, candidate model three, candidate model four) is constructed.

[0089] Model two in the candidate set is deleted, then model candidate scheme two (candidate model one, candidate model three, candidate model four) is constructed.

[0090] Model three in the candidate set is deleted, then model candidate scheme three (candidate model one, candidate model two, candidate model four) is constructed.

[0091] Model four in the candidate set is deleted, it is found that candidate model one, candidate model two and candidate model three cannot meet the task demand.

[0092] Delete alternative model two in alternative solution one, find that the combination of alternative model three and alternative model four cannot meet the user's demand, and for the same reason, delete alternative model three and alternative model four, then consider that the elements of alternative solution one cannot be deleted, and then determine solution one (alternative model two, alternative model three, alternative model four).

[0093] Delete alternative model one in alternative solution two, find that the combination of alternative model three and alternative model four cannot meet the user's demand, but after deleting alternative model three, the combination of alternative model one and alternative model four can meet the task demand, so the alternative solution four (alternative model one, alternative model four) is determined.

[0094] For the same reason, delete alternative model two in alternative solution two and alternative solution three, and the combination of alternative model one and alternative model four can meet the task demand, so the alternative solution of this step is the same as the alternative solution four constructed in the last step.

[0095] Delete alternative model one or alternative model four in alternative solution four, which cannot meet the task demand, so the solution two is (alternative model one, alternative model four).

[0096] Through the above examples, the solution one (alternative model two, alternative model three, alternative model four) and the solution two (alternative model one, alternative model four) are selected.

[0097] Next, the allocation strategy module also needs to determine how to complete the task in time and step by step. Specific examples are as follows: the current environment of the task provides a bandwidth of 100 bits / s, a CPU of 10, and a memory of 200G. And the solution one (alternative model two, alternative model three, alternative model four) and the solution two (alternative model one, alternative model four) have been obtained.

[0098] (1) First, calculate the resources of scheme one and scheme two, the bandwidth of scheme one is 140, CPU is 12, and memory is 160. The bandwidth of scheme two is 110, CPU is 8, and memory is 120. Compare the percentage of resources exceeding the existing environment in each scheme, select the largest limited index as the main index. For example, in scheme one, bandwidth: (140-100) / 100=40%, CPU: (12-10) / 10=20%, memory: (160-200) / 200=-20%. The main index of scheme one is bandwidth, which exceeds 40%. In scheme two, bandwidth: (110-100) / 100=10%, CPU: (8-10) / 10=-20%, memory: (120-200) / 200=-40%. The main index of scheme two is bandwidth, which exceeds 10%. Compare the excess of the main index in the scheme, and select the scheme with less excess. If the main index of scheme one exceeds more than the main index of scheme two, scheme two should be selected. If there is no excess index in the scheme, compare the scheme with the smallest main index, and start calculating all models.

[0099] (2) In order to describe the following model calculation process clearly, this embodiment selects scheme one as the description object. First, according to the proportion of the main index of the candidate model from large to small, the model is sorted. The main index of scheme one is bandwidth, which can be arranged as candidate model two, candidate model four, and candidate model three according to the bandwidth from large to small.

[0100] (3) According to the time to start, the task starts at 19:00-21:00, and the first time period 19:00-20:10 is calculated by the first candidate model two. After starting the first time period calculation, it is judged whether the remaining resources of the environment meet the next model requirement. If it meets the next model requirement, the next time period model calculation will be started. If it does not meet, the model calculation will be suspended, and the next model will be judged until the last model.

[0101] Therefore, when the candidate model two is calculated, the environment provides bandwidth 60, CPU 5, and memory 150. The resources of the second candidate model four are all less than the resources provided by the environment, so the calculation can be started at the same time.

[0102] The second time period 19:00-19:30 is calculated by the candidate model four. At this time, the resources provided by the environment do not meet the resource requirements of the candidate model three, and the candidate model three is the last model, so the time period is terminated.

[0103] When 19:30, the alternative model four is calculated, and it is determined whether the next model resource is sufficient at this time. If it is sufficient, the current time starts to calculate. For example, the required resource of the alternative model three is less than the environment provided resource, and the alternative model three is calculated in the second time period 19:30-20:10. If the resource is not sufficient, the model is arranged after the model in the previous time period. It is determined whether the resource of the next model is sufficient. Until the last model. For example, if the required resource of the alternative model three exceeds the environment provided resource at this time, the alternative model three is arranged to start calculating after the model in the previous time (i.e. in the first time period). For example, there is an alternative model five after the alternative model three in the scheme, and it is determined whether the alternative model five meets the current provided resource in the second time period, until the last model.

[0104] The above scheme is carried out in time and step by step, and the time period 1 and the time period 2 are calculated at 19:00. The first model in the time period 1 is model two, and the first model in the time period 2 is model four. Because the calculation time of model four is smaller than that of model two, model four will be calculated first in the time period 2. (That is, during the running time of model four 19:00-19:30, model four and model two run together.) Then it is determined whether model three can run in the time period 2. It is found that it cannot, and it can only start after model two is run, i.e. it is arranged in the time period 1.

[0105] For the credit calculation module, the output of the model in the current environment is evaluated according to the output parameter of the model in the previous time (historical time), and the credit index (credit label) of the model is constructed. The model credit label is represented by the following triple: Among them, represents the minimum resource required by the jth model in the L index of transmission, storage and calculation resource, represents the accuracy of the K index required by the calculation task of the model j, and the model j credit brief j is added to the label to form the model information label credit brief j is related to the input and output of the model j: it is assumed that the environment, input and output matrix of the model in the previous n times of calling are as follows:

[0106]

[0107] X represents the input matrix of the model, and Y represents the output matrix of the model. If the model only calculates part of the intermediate quantity when calling, the remaining input and output are filled with 0.

[0108] First, the data of X and Y is changed, and the element X jL is taken as an example, and the changed value is equal to the original value X jL minus the mean value E(X J). X jL =X jL -E(X J ).

[0109] Keep the input vector distance unchanged for each call and change X(n*L) to x new , Y changes to y new Because each model input represents the current network environment, and the input vector distance difference between different moments represents the network changes between the current moment and the previous moment, ensuring that the row vector distance in the input matrix remains unchanged can clearly predict the changes in the output matrix when the environment (input matrix) undergoes the same changes in the future.

[0110] Then the n-th distance matrix D can be calculated based on X, where i represents the i-th row and j represents the j-th column:

[0111] Among them D ij =dist(X i ,X j ).

[0112] Then assume Then keep D ij When unchanged:

[0113]

[0114] Perform eigenvalue decomposition on B, B = VAV T , then x new =A 1 / 2 V T .

[0115] Calculate the output y when calling input x(1*l): y=x new X -1 y new .

[0116] Compare the estimated value y with the task requirement Y to obtain the credit value:

[0117]

[0118] For the reward and punishment mechanism module, after selecting the model combination scheme with the least model requirements according to the allocation strategy module and calling the model combination to complete the task calculation, the reward and punishment mechanism module compares the actual model usage information with the model information label information to reward or punish the model credit.

[0119] In summary, this embodiment proposes a digital twin resource allocation architecture, which is based on four major modules: information collection, credit calculation, allocation strategy, and reward and punishment mechanism. Starting from the matching of model and user information, to predicting the model output in the current environment, and finally reasonably allocating the model based on storage, network, and computing resources, and finally establishing a reward and punishment mechanism for the model by verifying the performance of the model under the resource allocation scheme. This embodiment also proposes a model combination strategy based on user tasks and current environmental resources, as well as a model step-by-step time-sharing operation mechanism. By matching with user task requirements, a model combination method that meets environmental resources is selected, and then through different model combinations, a step-by-step time-sharing calculation method is adopted within the user-specified time, so that the models running at the same time do not exceed the provided transmission, storage, and computing resources, and complete the user task as quickly as possible.

[0120] Those skilled in the art will understand that all or part of the steps in the various methods of the embodiments can be completed by instructing related hardware through a program, and the program can be stored in a computer-readable storage medium, which may include: read-only memory (ROM), random access memory (RAM), a disk or an optical disk, etc.

[0121] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention. Any matters not described in detail in this specification constitute prior art known to those skilled in the art.

Claims

1. A digital twin resource allocation method, characterized in that: include: Analyze demand and establish clear demand indicators; Calculate the model's credit and estimate the model's output parameters; Select all model combinations that meet the requirements as alternatives; Select the best solution from the alternatives and implement it, and establish a feedback mechanism based on the actual use of the model; Calculating the model's credit and estimating the model's output parameters include: evaluating the model's output in the current environment based on the model's historical output parameters, and constructing a model's credit label; wherein constructing the model's credit label specifically includes: Obtain the resource requirements for transmission, storage, and computing required by the model. If the model does not require any of these resources, the resource is represented by 0. If any resource in the resource vector exceeds the resources currently available to the model, the model is considered unsuitable for the current network environment. Obtain the accuracy of the model when calculating multiple indicators of the current task. If the theoretical model accuracy does not meet the task requirements, it is considered that the model accuracy is limited in this network environment; When both resource requirements and accuracy meet the requirements of the current network environment, the model's credit label is calculated. The model's credit label is related to the model's input and output at the historical moment. The output and input relationship of the model at the historical moment is used to predict the model's output under the current network environment state. By comparing the difference between the model's predicted output and the task requirements, it is estimated whether the model can meet the task requirements, that is, to obtain the model's credit.

2. The digital twin resource allocation method according to claim 1, characterized in that: The analysis of needs and establishment of clear demand indicators specifically include: Obtain a description of the requirements, including transmission, storage, calculation, time, error rate, and credit. The descriptions of transmission, storage, and calculation indicators all include minimum and maximum values. The description of time indicators includes the start and end times of the run. The description of the error rate indicator includes the maximum acceptable calculation error of the demand parameters. The description of the credit indicator includes the minimum acceptable model credit.

3. The digital twin resource allocation method according to claim 1, characterized in that: When calculating the credit of the model, both the credit calculation of a single model and the credit calculation of a combination of multiple models are supported.

4. The digital twin resource allocation method according to claim 1, characterized in that: The selection of all model combination solutions that meet the requirements as alternative solutions specifically includes: According to the requirements, among all the models, as long as there is one model whose output meets the required requirements, the corresponding model will be used as an alternative model; Randomly sort all candidate models to obtain a set of candidate models; Starting from the first model, each time the model is deleted from the candidate model set, it is determined whether the remaining set still meets the task calculation result requirements. If it still meets the requirements, the deletion process is repeated, otherwise it is terminated; The final set that cannot be deleted is used as the alternatives, and the alternatives are arranged in ascending order based on the total amount of transmission, storage, and computing resources; Among the alternative plans, adjust the time resource planning for all alternative plans based on the required resources and time requirements.

5. The digital twin resource allocation method according to claim 4, characterized in that: The time resource planning adjustment for all alternative plans based on the required resources and time requirements specifically includes: If the total amount of the model's integrated resources in terms of transmission, storage, and calculation is greater than the required resources, then step-by-step and time-sharing calculations are used; otherwise, calculations are performed at the same time. When using step-by-step time-sharing calculation: arrange the models from high to low according to the centrality in terms of resources that exceed the demand; start with the first model as the first moment, add the second model from the first model as the sequence of that moment, and so on. If any resource exceeds the total after adding the next model, further judgment is made: if the resources are less than the total resources of the original sequence after removing the last model in the sequence and adding the next model, the last model is replaced with the next model to form a sequence, and the original last model is placed in the next moment.

6. The digital twin resource allocation method according to claim 5, characterized in that: When the model is arranged from high to low according to the centrality on the resources that exceed the demand, if there are multiple resources that exceed the demand, the one that exceeds the most is selected for arrangement.

7. The digital twin resource allocation method according to claim 1, characterized in that: The selecting the best solution from the alternative solutions to implement specifically includes: According to the allocation strategy, the model resources occupied and the model credit in each solution are compared, and the solution with the least resource occupation, the least computing time and the highest credit is selected as the optimal solution to be implemented.

8. A digital twin resource allocation architecture, characterized in that: It includes information collection module, credit calculation module, allocation strategy module and penalty mechanism module, among which: The information collection module is used to construct a model query table LFM, which includes task requirements, resource start and end time, task indicators and indicator accuracy; The credit calculation module is used to determine which model outputs meet the task indicators after receiving the LFM, and determine whether multiple combinations of models can meet the task requirements. If so, the credit calculation module extracts various information of the models that meet the requirements from the data pool to construct an information label and pass it to the allocation strategy module; The allocation strategy module is used to select a combination of multiple models based on the model information in the information tag that satisfies the credit calculation construction, determine how to complete the task in a time-sharing and step-by-step manner under resource constraints, and then send the task and model information to the probability model and simulation model to call the model combination to complete the task calculation; The penalty mechanism module is used to compare the actual model usage information with the model information label information to reward or punish the model credit; The credit calculation module calculates the credit of the model and estimates the output parameters of the model, specifically including: evaluating the output of the model in the current environment based on the output parameters of the model at historical moments, and constructing the credit label of the model; The credit labels for building the model specifically include: Obtain the resource requirements for transmission, storage, and computing required by the model. If the model does not require any of these resources, the resource is represented by 0. If any resource in the resource vector exceeds the resources currently available to the model, the model is considered unsuitable for the current network environment. Obtain the accuracy of the model when calculating multiple indicators of the current task. If the theoretical model accuracy does not meet the task requirements, it is considered that the model accuracy is limited in this network environment; When both resource requirements and accuracy meet the requirements of the current network environment, the model's credit label is calculated. The model's credit label is related to the model's input and output at the historical moment. The output and input relationship of the model at the historical moment is used to predict the model's output under the current network environment state. By comparing the difference between the model's predicted output and the task requirements, it is estimated whether the model can meet the task requirements, that is, to obtain the model's credit.

Images