Edge network digital twin application migration method based on elastic prediction

Abstract

The invention discloses an edge network digital twin application migration method based on elastic prediction, and belongs to the technical field of network resource management. The method is used for digital twinning auxiliary task execution under end-edge-cloud collaboration, and the method is based on elastic movement prediction and actively migrates digital twinning to adapt to future position changes of corresponding physical entities. In order to minimize task response delay and meet energy consumption constraints, the method provided by the invention comprises the following steps of: adopting a dual-time-scale online joint optimization method to decide and predict the elastic uploading frequency of network input information, a digital twin migration decision and calculation and communication resource allocation.

Description

Technical Field

[0001] The present invention belongs to the technical field of network resource management, specifically relates to the migration of digital twin models at the network edge, and particularly relates to a method for migrating digital twin applications in edge networks based on elastic prediction. Background Art

[0002] Digital twin has been regarded as a key technology to support 6G Internet-of-Things applications in recent years. It is a forward-looking paradigm for constructing virtual representations of physical entities. The digital twins of physical objects (such as mobile users and vehicles) are established based on their historical data and real-time states. Through simulation and behavior analysis in the digital space, digital twins can achieve close monitoring of physical entities and provide intelligent decision-making support for actual services. To meet the high service experience requirements of physical twins, their corresponding digital twins need to be well constructed and managed to provide low-latency and energy-saving task execution services. These requirements have promoted the application of end-edge-cloud collaboration. Digital twins can be deployed at the network edge and supported by the cloud center and terminal physical entities. Although some existing studies have considered the deployment of digital twins at the network edge, there are still many problems that have not been solved, mainly including the following aspects:

[0003] 1) Different from fixed industrial devices, physical twins move continuously, which may lead to instability in the connection between physical twins and digital twins. This means that digital twins must be dynamically migrated between multiple edge servers to provide seamless task execution services for physical twins.

[0004] 2) Different from the limited number of and shareable popular service applications, the system needs to deploy a large number of exclusive DTs for all physical entities to meet personalized service requirements. In this scenario, each edge server must maintain the continuous operation of multiple digital twin models simultaneously, which may lead to overloading and thus bring extremely high task response latency and even service interruption. Therefore, it is necessary to actively migrate digital twin models by predicting the future movement trajectories of PTs.

[0005] However, if all the above factors are considered, it will be difficult to implement the digital twin service system for this edge network, for the following three reasons:

[0006] a. Considering that the physical entity state information for mobility prediction is private and has a large amount of data, which will cause additional upload energy consumption, so this data should be collected and uploaded at a lower frequency. In contrast, due to the continuous movement of users and the continuous generation of tasks, digital twin model migration and the corresponding communication and computing resource allocation need to be optimized at a higher frequency. However, the communication channel is shared, so various decisions at different time scales need to be optimized collaboratively.

[0007] b. The frequency of uploading the physical entity status data affects the task response latency and energy consumption: the shorter the uploading interval, the smaller the task execution latency, but the higher the energy consumption. While the longer the uploading interval, the larger the task execution latency, but the correspondingly lower the energy consumption. This means that the uploading frequency should be adaptively and flexibly adjusted based on the feedback of latency and energy consumption.

[0008] c. Due to the extremely fast movement speed of physical entities and the large number of digital twin migration requirements, traditional passive migration strategies may lead to extremely high task response latency and even service interruption. Therefore, the migration of DTs should be actively carried out by constructing a mobility prediction model. Summary of the Invention

[0009] Object of the Invention: Aiming at the deficiencies of the existing technologies in the digital twin migration of edge networks mentioned in the above background technology, the present invention provides a digital twin application migration method for edge networks based on elastic prediction.

[0010] Technical Solution: A digital twin application migration method for edge networks based on elastic prediction, the method is based on mobility prediction, faces physical entity-digital twin pairs, aims to minimize task response latency, and dynamically and real-timely actively migrates the digital twin models in the system. This method determines the elastic uploading frequency of the prediction network input information, digital twin migration decisions, and computing and communication resource allocation through online joint optimization on a dual-time scale;

[0011] The method includes the following steps:

[0012] (1) For the edge computing system, construct a digital twin migration model based on mobility prediction under the end-edge-cloud network architecture, including a number of physical entity-digital twin pairs, edge servers, and a central cloud;

[0013] (2) Calculate the total system energy consumption at large and small time scales respectively in each time step. The total system energy consumption at the large time scale includes the energy consumption of each physical entity uploading the information required by the prediction network, and the total system energy consumption at the small time scale includes the migration energy consumption, task transmission energy consumption, and task processing energy consumption of each digital twin model;

[0014] (3) Optimize the digital twin migration problem in the model, and the mathematical representation of this optimization process is:

[0015]

[0016] where, T reap (t) represents the calculated task response latency, and the calculated task response latency includes task transmission latency and task processing latency, τ k represents the large time scale decision, Represents the small time-scale decisions determined within each time step,

[0017] The constraints are as follows:

[0018]

[0019]

[0020] Among them, a i,m (t) represents the migration decision, and u i (t) represents the communication resource allocation decision, and f i (t) represents the computing resource allocation decision. C is the total system energy consumption limit value, and E sts (t) is the system energy consumption at the small time scale. A large time-scale energy consumption will occur only when t = t (k) . E lts (t (k) ) is the system energy consumption at the large time scale;

[0021] (4) Solve the optimization problem in step (3), including:

[0022] Define an energy consumption deficit queue Q(t) to describe the degree to which the system energy consumption deviates from the long-term average energy consumption C / T;

[0023] Use the Lyapunov optimization method to decompose the original problem into sub-problems for each time step and solve them;

[0024] Adopt a prediction method based on gated recurrent units to predict the positions of physical entities in each time step.

[0025] For the large time-scale decision τ k , use the upper confidence bound algorithm to learn the optimal decision;

[0026] For the small time-scale decision S(t), use the method based on alternating minimization to solve.

[0027] Furthermore, in step (2), the task transmission delay is expressed as follows:

[0028]

[0029] The task processing delay is expressed as follows:

[0030]

[0031] The total system energy consumption at the large time scale and the small time scale are calculated as follows:

[0032]

[0033] Among them, calculate the energy consumption of uploading status information from the physical entity to the central cloud The expression is as follows:

[0034]

[0035] This energy consumption is divided into the energy consumption of uploading status information from the physical entity to the edge server and the energy consumption of uploading from the edge server to the central cloud Their expressions are respectively and

[0036] Task upload energy consumption The expression is as follows:

[0037]

[0038] Task processing energy consumption The expression is as follows:

[0039]

[0040] Digital twin migration energy consumption The expression is as follows:

[0041]

[0042] In the formula, λ i (t) is the amount of task data generated by the physical entity within time t, r i,m (t) represents the transmission rate between the physical entity and the edge server, C m represents the number of CPU cycles of the edge server, F m represents the CPU processing speed, ρ m represents the effective switching capacitance of the edge server, φ represents the energy consumption of transmitting unit data per unit distance, D i represents the size of the digital twin model to be migrated.

[0043] Furthermore, the calculation expression for updating the energy consumption deficit queue Q(t) at each time step is as follows:

[0044]

[0045] The update expression for the hidden unit state of the gated recurrent unit prediction network in this method is as follows:

[0046] r t =σ(W ir x t +b ir +W hr ht-1 +b hr ),

[0047] z t = σ(W iz x t +b iz +W hz h t-1 +b hz ),

[0048] n t = tanh(W in x t +b in +r t (W hn h t-1 +b hn ))

[0049] h t = (1 - z t )n t + z t h t-1 ,

[0050] y t = W o h t +b o ,

[0051] where W and b are the training parameters of the model. W ir , W iz , W in represent the input to the reset gate, update gate, and candidate hidden state matrices, and W hr , W hz , W hn represent the weight matrices from the previous hidden state to the reset gate, update gate, and candidate hidden state; b ir , b hr , b iz represent the bias terms corresponding to the input part, and b hz , b in , b hn represent the bias terms corresponding to the hidden state part; W o is the weight matrix from the hidden state to the output, and b o is the bias term of the output layer. r t , z t , n t are the reset gate, update gate, and new gate at time step t respectively, h t is the hidden state at time step t, x t and y tare input and output variables, and σ(·) is the sigmoid function. When there is no state information as input, the prediction network can generate the output of the current time step through autoregression;

[0052] In this method, learning the optimal decision at a large time scale based on the upper confidence bound algorithm specifically includes:

[0053] The sub-problem expression of the optimal state information upload frequency is as follows:

[0054]

[0055] Constraint conditions:

[0056] Then, the optimal decision is selected according to the upper confidence bound of each feasible arm;

[0057] The optimal decision is expressed as where T τ is the number of times τ is selected, r(τ) is the reward function for selecting τ, and r(τ) = -T resp (τ) + p(τ), and p(τ) is the penalty term for violating the energy consumption constraint,

[0058] In this method, the decision-making at a small time scale is solved based on the method of alternating minimization Specifically, it includes:

[0059]

[0060] where P 4-1 is a linear integer programming problem, P 4-2-1 and P 4-2-2 are convex functions and are solved using a solver; V represents the Lyapunov control parameter, and Q(t) represents the queue backlog at time t.

[0061] Beneficial effects: Compared with the prior art, the substantial progress and remarkable effects of the present invention are as follows:

[0062] (1) The present invention takes into account the one-to-one characteristics of the digital twin and the physical entity and the mobility of the physical entity, and realizes the active migration of the digital twin model on the edge server to ensure the low latency and reliability of the physical entity task processing.

[0063] (2) The method of the present invention comprehensively considers the influencing factors of the delay and energy consumption of task processing, including the frequency of uploading the input information of the mobile prediction network for active migration, the specific migration decision, and the allocation of computing and communication resources, and jointly optimizes all the above decisions.

[0064] (3) By analyzing the characteristics of the problems studied in the present invention, a dual-time-scale optimization algorithm is designed to solve the optimization decision. The Lyapunov optimization framework is adopted to decompose the original problem and solve it at each time step. The upper confidence bound algorithm and the alternating minimization method are used to solve the decisions on the two time scales respectively, and finally the minimization of the task processing delay and energy consumption is achieved. Description of the Drawings

[0065] Figure 1 It is a model diagram of the digital twin migration system under the end-edge-cloud architecture in the embodiment;

[0066] Figure 2 It is a topological diagram of the edge server network in a certain area in the embodiment;

[0067] Figure 3 shows a comparison diagram of the long-term average task delay and system energy consumption between the embodiment and the prior art under different numbers of physical entities. Detailed Embodiment

[0068] In order to elaborate in detail the technical solution disclosed by the present invention, the following further elaboration is made in combination with the drawings in the specification and specific embodiments.

[0069] Considering the problem of unstable connection between the physical entity and the digital twin model caused by the high-speed movement of the physical entity in the scenario of digital twin-assisted task processing, and the problem of tight resource allocation caused by the need for a dedicated digital twin model for each physical entity, the present invention provides a method for actively migrating digital twin models at the edge. Specifically, the method deploys a dedicated digital twin model for each physical entity at the network edge to process its complex tasks. Since the physical entity has mobility, its corresponding digital twin will migrate between edge servers according to the future movement trajectory of the physical entity, which is predicted by the mobility prediction model constructed by the cloud center. To ensure the operation of the prediction model, the physical entity needs to upload its own status information (such as movement speed, direction, etc.) as the model input.

[0070] The present invention provides a method for migrating digital twin applications in an edge network based on elastic prediction. Based on the method described in the present invention, we propose to adopt a dual-time-scale online optimization mechanism to minimize the average task response delay in the long term under the premise of meeting strict energy consumption constraints, while considering system uncertainty (i.e., the mobility of physical entities). Different from the prior art, this method considers that the upload frequency of the physical entity status data required for predicting network input in the large time scale is elastically variable. Specifically, the method described in the present invention includes constructing a dual-time-scale joint optimization scheme to determine the data upload frequency in the large time scale and the migration decision, the computing and communication resource allocation decision of the edge server in the small time scale, so as to ensure the reliability of task processing and the minimization of the long-term average task processing delay and the total system energy consumption.

[0071] Further, in combination with Figure 1 , the implementation process of the method described in the present invention is elaborated in detail as follows:

[0072] Step 1: Build an edge server network capable of deploying and migrating digital twin models, and construct a mathematical model of the system.

[0073] This step mathematically analyzes the actual problems solved by the edge network digital twin application migration method based on elastic prediction.

[0074] According to the actual application scenario, the digital twin-assisted task execution system based on edge computing consists of a cloud center, M edge servers, and I mobile physical entities and their digital twin models. The cloud center serves as the central controller and uses a model based on gated recurrent units to achieve the active migration of digital twin models. Each physical entity generates a series of complex tasks, which need to be executed by the dedicated digital twin models deployed at the edge. Due to the complex movement patterns of physical entities, their associated digital twin models should be actively migrated between different edge servers according to the movement prediction model constructed by the cloud center. At the same time, the physical entities should also switch the edge servers they connect to in order to access the latest migrated location of their digital twin models. Each edge server can provide digital twin-assisted task processing services for multiple physical entities simultaneously, and its communication and computing resources are shared among these digital twin models. Once the digital twin model migration is completed, the physical entity can offload the task to its corresponding digital twin model for execution.

[0075] In the above digital twin-assisted task execution system, the task transmission delay is expressed as follows:

[0076]

[0077] The task processing delay is expressed as follows:

[0078]

[0079] where λ i (t) is the amount of task data generated by the physical entity within time t.

[0080] The total energy consumption of the system on large and small time scales is calculated as follows:

[0081]

[0082] Among them, the energy consumption for calculating the upload of status information from the physical entity to the central cloud is expressed as:

[0083]

[0084] The energy consumption is divided into the energy consumption for uploading the status information from the physical entity to the edge server and the energy consumption for further uploading from the edge server to the central cloud Their expressions are respectively and where J i (t) is the amount of input data used for mobile prediction; p i is the transmission power of the physical entity. r m and p m are the uplink transmission rate and the unit transmission power through the optical fiber link channel respectively.

[0085] Task upload energy consumption The expression is:

[0086]

[0087] Task processing energy consumption The expression is:

[0088]

[0089] where C m is the number of CPU cycles of the edge server, and ρ m represents the effective switching capacitance of the edge server.

[0090] Digital twin migration energy consumption The expression is:

[0091]

[0092] where φ represents the energy consumption for transmitting unit data per unit distance, and D i is the size of the digital twin model to be migrated.

[0093] Next, a mathematical model is established for the problem of minimizing the task processing delay of the system under the conditions of meeting the edge server resource allocation conditions and the maximum system energy consumption constraint. The mathematical expression is:

[0094]

[0095] Constraint conditions:

[0096]

[0097] where the total system energy consumption on large time scale and small time scale is calculated as follows:

[0098]

[0099] Step 2: Problem transformation.

[0100] Based on the original optimization problem in Step 1, first define an energy consumption deficit queue to reflect the deviation of the total system energy consumption from the long-term average energy consumption budget C / T. The evolution of this queue over time can be expressed as:

[0101]

[0102] Traditional Lyapunov optimization methods require the use of fixed constraints to ensure the boundedness of the performance gap. However, since the state information of physical entities is only uploaded at the beginning of each large time step, the total system energy consumption will mutate at that moment. To ensure the stability of the energy consumption deficit queue, the energy consumption for uploading state information is evenly distributed over each small time step within the large time step τ k Therefore, for any small time step t within the large time step τ k the evolution of the queue can be corrected to:

[0103]

[0104] Next, according to the drift-plus-penalty term expression of the Lyapunov optimization framework, we can obtain

[0105]

[0106] where Therefore, the original problem can be decomposed into a series of sub-problems, where the right-hand side of the above equation is minimized in each large time step k:

[0107]

[0108] For the decomposed sub-problems, a prediction network based on gated recurrent units is used to predict the positions of physical entities in each small time step. The update expressions for the hidden unit states are as follows:

[0109] r t = σ(W ir x t + b ir + W hr h t-1 + b hr ),

[0110] z t = σ(W iz x t + b iz + W hz h t-1 + b hz ),

[0111] n t = tanh(W in x t + b in + r t (W hn h t-1 + b hn ))

[0112] h t = (1 - z t )n t + z t h t-1 ,

[0113] y t = W o h t + b o ,

[0114] where W and b are the training parameters of the model, r t , z t , n t are the reset gate, update gate, and new gate at time step t respectively, h t is the hidden state at time step t, x t and y t are the input and output variables, and σ(·) is the sigmoid function. When there is no state information as input, the prediction network can generate the output of the current time step through autoregression.

[0115] Furthermore, in order to obtain the upload frequency of the physical entity state information as the input of the prediction network at each large time step, the sub-problem expression on the large time scale is as follows:

[0116]

[0117] Constraint:

[0118] Select the optimal decision according to the upper confidence bound of each feasible arm, that is where T τ is the number of times τ is selected, r(τ) = -T resp (τ) + p(τ) is the reward function for selecting τ, is the penalty term for violating the energy consumption constraint.

[0119] Finally, for the solution of each small time step decision, there is a sub-problem:

[0120]

[0121] where P4-1 is a linear integer programming problem, P 4-2-1 and P 4-2-2 are convex functions and can both be solved by existing solvers (such as Gurobi, etc.). By iteratively solving through the alternating minimization method, the small-time scale decision of the optimization problem can ultimately be obtained.

[0122] As Figure 2 shown, we consider a digital twin-assisted task processing system on the edge network within a 15 km × 18 km geographical area, which includes M = 20 edge servers, each located at the center of a hexagonal grid. Multiple physical entities (including vehicles and pedestrians) move within this area, continuously generating complex tasks that need to be uploaded to the corresponding digital twin for processing.

[0123] Combined with Figure 3, the superior performance of the digital twin application migration method based on elastic prediction for edge networks described in the present invention is verified compared with existing technical methods. From Figure 3(a), we can see that under different numbers of physical entities, this method has the best performance in terms of task processing latency. From Figure 3(b), it can be found that under different numbers of physical entities, the total system energy consumption of this method is also the lowest.

Claims

1. A method for migrating digital twin applications in an edge network based on elastic prediction. The method is based on mobility prediction, targets physical entity-digital twin pairs, and dynamically and real-timely proactively migrates digital twin models in the system with the goal of minimizing task response latency. It is characterized in that: The method determines the elastic upload frequency of the input information of the prediction network, the digital twin migration decision, and the allocation of computing and communication resources through online joint optimization of double time scales; The method includes the following steps: (1) For the edge computing system, a digital twin migration model based on mobility prediction is constructed under the end-edge-cloud network architecture, including several physical entity-digital twin pairs, edge servers, and a central cloud; (2) Calculate the total system energy consumption at large and small time scales respectively in each time step. The total system energy consumption at large time scales includes the energy consumption of each physical entity uploading the information required by the prediction network, and the total system energy consumption at small time scales includes the migration energy consumption, task transmission energy consumption, and task processing energy consumption of each digital twin model; (3) Optimize the digital twin migration problem in the model. The mathematical representation of this optimization process is: Among them, T resp (t) represents the computing task response delay, and the computing task response delay includes task transmission delay and task processing delay, τ k represents the large time scale decision, represents the small time scale decision determined within each time step, The constraints are as follows: Among them, a i，m (t) represents the migration decision, u i (t) represents the communication resource allocation decision, f i (t) represents the computing resource allocation decision, c is the total system energy consumption limit value, E sts (t) is the system energy consumption at a small time scale, and a large time scale energy consumption occurs only when t = t (k) ; E lts (t( k) ) is the system energy consumption at a large time scale; (4) Solve the optimization problem in step (3), including: Define the energy consumption deficit queue Q(t) to describe the degree to which the system energy consumption deviates from the long-term average energy consumption C / T; Use the Lyapunov optimization method to decompose the original problem into sub-problems at multiple time steps for solution; Adopt a prediction method based on a gated recurrent unit to predict the position of physical entities in each time step. For decisions τ on large time scales k , the upper confidence bound algorithm is used to learn the optimal decision; For the decision S(t) at small time scales, use the method based on alternating minimization to solve.

2. The method for migrating digital twin applications of an edge network based on elastic prediction according to claim 1, wherein: In step (2), the task transmission delay has the following expression: Task processing delay The expression is as follows: The total system energy consumption at large and small time scales is calculated as follows respectively: Among them, the energy consumption for calculating the upload of status information from physical entities to the central cloud is expressed as: The energy consumption is divided into the energy consumption for uploading status information from physical entities to the edge server and the energy consumption for further uploading from the edge server to the central cloud Their expressions are respectively and The expression for the energy consumption of task uploading is as follows: Task processing energy consumption The expression is as follows: Digital twin migration energy consumption The expression is as follows: where λ i (t) is the amount of task data generated by the physical entity within time t, and r i，m (t) represents the transmission rate between the physical entity and the edge server station, and C m represents the number of CPU cycles of the edge server, and F m represents the CPU processing speed, and ρ m represents the effective switching capacitance of the edge server, φ represents the energy consumption for transmitting unit data per unit distance, and D i represents the size of the digital twin model that needs to be migrated.

3. The method for migrating digital twin applications of an edge network based on elastic prediction according to claim 1, wherein: The calculation expression for updating the energy consumption deficit queue Q(t) in each time step is as follows:

4. The method for migrating digital twin applications in an edge network based on elastic prediction according to claim 1, wherein: The update expression for the hidden unit state of the gated recurrent unit prediction network is as follows: r t = σ(W ir x t + b ir + W hr h t-1 + b hr ), z t = σ(W iz x t + b iz + W hz h t-1 + b hz ) n t = tanh(W in x t + b in + r t (W hn h t-1 + b hn )) h t = (1 - z t )n t + z t h t-1 , y t = W o h t + b o , Among them, W and b are the training parameters of the model. W ir , W iz , W in represent the input to the reset gate, update gate, and candidate hidden state matrix. W hr , W hz , W hn represent the weight matrices from the previous hidden state to the reset gate, update gate, and candidate hidden state; b ir , b hr , b iz represent the bias terms corresponding to the input part. b hz , b in , b hn represent the bias terms corresponding to the hidden state part; W o is the weight matrix from the hidden state to the output, and b o is the bias term of the output layer; r t , z t , n t are the reset gate, update gate, and new gate at time step t respectively; h t is the hidden state at time step t, x t and y t are the input and output variables, and σ(·) is the sigmoid function. When there is no state information as input, this prediction network can generate the output of the current time step through autoregression.

5. The method for migrating digital twin applications in an edge network based on elastic prediction according to claim 1, wherein: Learning the optimal decision at large time scales based on the upper confidence bound algorithm specifically includes: The sub-problem expression for uploading the optimal state information frequency is as follows: Constraints: Then select the optimal decision according to the upper confidence bound of each feasible arm; The optimal decision is expressed as where T τ is the number of times τ is selected, r(τ) is the reward function for selecting τ, and r(τ) = -T retp (τ) + p(τ), where p(τ) is the penalty term for violating the energy consumption constraint 6. The edge network digital twin application migration method based on elastic prediction according to claim 1, characterized in that: Method based on alternating minimization for solving decisions at small time scales Specifically include: where P 4-1 is a linear integer programming problem, P 4-2-1 and P 4-2-2 are convex functions, which are solved using a solver; V represents the Lyapunov control parameter, and Q(T) represents the queue backlog at time t.

Images