A 5G Distribution Network Resource Scheduling Method Assisted by Digital Twin

Abstract

The present invention relates to a 5G distribution network resource scheduling method assisted by digital twin, belonging to the field of communication technology. The object of the present invention is to provide a 5G distribution network resource scheduling method assisted by digital twin. By jointly optimizing terminal scheduling, terminal computing resource allocation, power control, and server computing resource allocation, while ensuring the long-term constraints of terminal energy consumption and access priority, the cumulative iterative delay of DT construction and terminal energy consumption are reduced, and the DT accuracy and security are improved.

Description

Technical Field

[0001] The present invention relates to a 5G distribution network resource scheduling method assisted by digital twin, belonging to the field of communication technology. Background Art

[0002] At present, a large number of Internet of Things terminals such as temperature and current sensors are deployed in the distribution network to provide real-time monitoring services, which puts forward higher requirements for low-latency data transmission and processing. Traditional wired and wireless communication methods, such as optical fibers, LTE-230, etc., cannot solve the communication and computing problems of the last kilometer of the distribution network due to high deployment costs, insufficient wireless spectrum resources, and insufficient edge intelligence. 5G edge computing provides a feasible solution by integrating advanced communication and computing technologies at the network edge. In order to further improve resource utilization efficiency and meet service quality requirements, dynamic intelligent scheduling of multi-dimensional resources such as communication, computing, and energy is required.

[0003] Digital twin (DT) can help the distribution network supported by 5G edge computing achieve multi-dimensional resource scheduling from the perspective of the digital world. DT collects real-time status from a large number of terminals and establishes digital representations of all physical entities, thereby realizing precise decision-making guidance for multi-dimensional resource scheduling. However, when constructing DT and using it to assist resource scheduling, some key challenges need to be addressed:

[0004] (1) Strict requirements of DT for latency, accuracy, and security: DT needs to be updated with low latency to maintain real-time consistency with physical entities. At the same time, in order to improve the accuracy of resource scheduling, the loss function reflecting the deviation between DT and the actual value must be minimized. On the other hand, the data collected from a large number of terminals for constructing DT also faces privacy issues and security threats. For example, abnormal model parameters uploaded by some malicious terminals will seriously reduce the accuracy of DT and cause various security problems.

[0005] (2) Joint optimization of multi-dimensional resources

[0006] The joint optimization problem of multi-dimensional resource scheduling is a very important problem because multi-dimensional resource scheduling is a coupling between different entities, dimensions, and processes. For example, terminal power control and computing resource allocation strategies are coupled, and server edge aggregation and abnormal model identification are coupled.

[0007] (3) Access priority and long-term energy consumption constraints: Considering the importance of services and the limited battery capacity of terminals, long-term access priority and energy consumption constraints need to be established to ensure the access priority of important terminals and reduce energy consumption. However, long-term constraints are coupled with short-term resource scheduling optimization.

[0008] In view of the above-mentioned deficiencies, the present invention aims to create a digital twin-assisted 5G distribution network resource scheduling method to make it more valuable in industrial applications. Summary of the Invention

[0009] To solve the above technical problems, the object of the present invention is to provide a digital twin-assisted 5G distribution network resource scheduling method. By jointly optimizing terminal scheduling, terminal computing resource allocation, power control, and server computing resource allocation, under the premise of ensuring long-term constraints on terminal energy consumption and access priority, the cumulative iterative delay of DT construction and terminal energy consumption are reduced, and the accuracy and security of DT are improved.

[0010] A digital twin-assisted 5G distribution network resource scheduling method of the present invention has the following specific scheduling steps:

[0011] S1. Build a system model, specifically including:

[0012] 1.1. Local training model;

[0013] 1.2. Transmission model;

[0014] 1.3. Packet error rate model;

[0015] 1.4. Edge aggregation and anomaly model detection model;

[0016] 1.5. Energy consumption and delay model;

[0017] S2. Problem modeling and transformation, specifically including:

[0018] 2.1. Access priority constraint;

[0019] 2.2. Energy consumption constraint;

[0020] 2.3. Problem modeling;

[0021] 2.4. Problem transformation;

[0022] S3. Secure and delay-aware digital twin-assisted resource scheduling method, specifically including:

[0023] 3.1. Joint optimization sub-problem of terminal access scheduling, terminal computing resource allocation, and power control;

[0024] 3.2. Server computing resource allocation sub-problem.

[0025] Furthermore, the system model construction in S1 is based on a federated learning-based 5G edge computing distribution network DT architecture, which mainly includes three layers, namely the terminal layer, the edge layer, and the DT layer.

[0026] Further, in the local training model scenario of S1, there are I power IoT terminals, and the set is represented as Y = {u1, u2,..., u I}, the total time period is divided into T iterations, and the set is represented as Τ = {1, 2,..., T}. The energy consumption and delay of local computing can be calculated as:

[0027]

[0028] where α is the energy consumption coefficient, D i represents the size of the local dataset Δ i of u i .

[0029] Define and to represent the input and target output of a sample in the dataset Δ i . The single-sample loss function is used to characterize the deviation between the estimated output and the target output. Therefore, the loss function of u i can be calculated as:

[0030]

[0031] The loss function can reflect the accuracy of the local model and is used to guide the update of the parameter ω i (t), that is:

[0032]

[0033] where η is the learning rate.

[0034] Further, the transmission rate of the transmission model in S1 can be expressed as:

[0035]

[0036] The transmission energy consumption and delay can be expressed as:

[0037]

[0038] Further, the packet error rate of the packet error rate model in S1 can be expressed as;

[0039]

[0040] where k is a constant related to the coding gain. Define the binary indicator variable a i (t) ∈ {0, 1}, a i (t) = 1 indicates that there is no packet error in the transmission of the terminal u i , otherwise a i (t) = 0, a i(t) can be calculated as:

[0041]

[0042] Furthermore, for the edge aggregation and anomaly model detection model in S1: The global model aggregation formula is:

[0043]

[0044] The edge aggregation delay is calculated as:

[0045]

[0046] The loss function of the global model is expressed as:

[0047]

[0048] Where represents the total number of samples of all selected terminals;

[0049] For the anomaly model detection model of the edge aggregation and anomaly model detection model in S1:

[0050] The anomaly model detection model delay can be calculated as:

[0051]

[0052] Where ξ0 represents the number of CPU cycles required for the edge server to process a single sample. When the uploaded anomaly model is greater than the ratio υ, that is: i When the uploaded anomaly model is greater than the ratio υ, that is:

[0053]

[0054] Determine u i as a malicious terminal and remove it from the set Σ(t + 1).

[0055] Furthermore, for u i in the energy consumption and delay model in S1, the total energy consumption is the sum of the local computing energy consumption and the transmission energy consumption, and is calculated as:

[0056]

[0057] The total delay of a single iteration includes the maximum delay of the sum of the local computing delay and the parameter transmission delay among all terminals, the larger delay between the edge aggregation delay and the AMR delay, and the global model broadcast delay τ b , and is expressed as:

[0058]

[0059] Where,

[0060] Furthermore, in the problem modeling and transformation in S2, access priority constraints are incorporated. The long-term access priority constraint is defined as:

[0061]

[0062] where e i is the minimum proportion requirement for the number of times of selecting u i for local training and parameter transmission.

[0063] Furthermore, considering the limited terminal battery capacity in the energy consumption constraint in S2, the long-term energy consumption constraint is defined as

[0064]

[0065] where, E i,max is the energy budget for u i .

[0066] Furthermore, the optimization objective of the problem modeling in S2 is defined as minimizing the weighted sum of the DT loss function and the total iteration delay through joint optimization of computing resource allocation, power control, and access scheduling under the long-term constraints of access priority and energy consumption, expressed as:

[0067]

[0068]

[0069]

[0070]

[0071]

[0072]

[0073]

[0074]

[0075]

[0076] Based on the relationship between the packet error rate and the DT loss function, P1 can be rewritten as:

[0077]

[0078] s.t.C1~C8.

[0079] where,

[0080] Furthermore, in S2, the problem transformation uses the concept of virtual queues to decouple long-term constraints and short-term resource scheduling optimization. C7 and C8 can be transformed into queue stability constraints, expressed as:

[0081] F i (t + 1) = max{F i (t) + e i -x i (t), 0},

[0082]

[0083] If N i (t) and Y i (t) are average rate stable, then C7 and C8 are automatically satisfied, and P2 can be transformed into:

[0084]

[0085] s.t. C1~C6,

[0086] C9: F i (t) and Y i (t) are average rate stable

[0087] Define the vector Θ(t) = [F i (t), Y i (t)], and the Lyapunov function can be expressed as:

[0088]

[0089] The Lyapunov drift is defined as the conditional expectation change of the Lyapunov function in two consecutive time slots, expressed as:

[0090] ΔL(Θ(t)) = E[L(Θ(t + 1)) - L(Θ(t))|Θ(t)]

[0091] To achieve the minimization of the weighted sum of the global loss function and the total iteration delay under the queue stability constraints, the drift-plus-penalty is defined as:

[0092] Δ V L(Θ(t)) = ΔL(Θ(t)) + VE[Φ(t)|Θ(t)]

[0093] where V is the weight of the P3 optimization objective. Therefore, P3 can be transformed into the problem of minimizing the upper bound of Δ V L(Θ(t)) under the C1~C6 constraints, and the upper bound of Δ V L(Θ(t)) is expressed as:

[0094]

[0095] Furthermore, the S3 secure and latency-aware digital twin-assisted resource scheduling algorithm solves the joint optimization problem P3, which can be decomposed into two sub-problems, namely SP1: the joint optimization sub-problem of terminal access scheduling, terminal computing resource allocation, and power control; SP2: the server computing resource allocation sub-problem;

[0096] Among them, SP1 is expressed as:

[0097]

[0098] SP2 is expressed as:

[0099]

[0100] s.t.C5,C6.

[0101] Since τ a (t) and τ d (t) are inversely proportional to and respectively, and Therefore, when τ a (t) = τ d (t), SP2 reaches the optimum, and the optimal server computing resource allocation can be expressed as:

[0102]

[0103] With the above solution, the present invention has at least the following advantages:

[0104] The present invention proposes a digital twin-assisted 5G distribution network resource scheduling method. By jointly optimizing terminal scheduling, terminal computing resource allocation, power control, and server computing resource allocation, it reduces the cumulative iterative latency of DT construction and terminal energy consumption while ensuring the long-term constraints of terminal energy consumption and access priority, and improves the accuracy and security of DT.

[0105] The above description is only an overview of the technical solution of the present invention. In order to understand the technical means of the present invention more clearly and implement it according to the content of the specification, the following describes the preferred embodiments of the present invention in detail in conjunction with the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0106] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for the embodiments. It should be understood that the following drawings only show a certain embodiment of the present invention, so it should not be regarded as a limitation of the scope. For those of ordinary skill in the art, other related drawings can also be obtained based on these drawings without creative efforts.

[0107] Figure 1 is the DT architecture of the 5G edge computing distribution network based on federated learning of the present invention;

[0108] Figure 2 is the framework diagram of SAINT of the present invention;

[0109] Figure 3 is the DT loss function diagram in the simulation experiment of the present invention;

[0110] Figure 4 is the cumulative iteration delay diagram in the simulation experiment of the present invention;

[0111] Figure 5 is the cumulative energy consumption diagram in the simulation experiment of the present invention. Specific Embodiments

[0112] The following combines the accompanying drawings and embodiments to further describe in detail the specific embodiments of the present invention. The following embodiments are used to illustrate the present invention, but are not used to limit the scope of the present invention.

[0113] A 5G distribution network resource scheduling method assisted by digital twin according to a preferred embodiment of the present invention. The entire technical solution includes three steps:

[0114] 1. Construct a system model

[0115] The DT architecture of the 5G edge computing distribution network based on federated learning is as Figure 1 shown. The purpose is to construct a digital model M of the physical network through federated learning. The architecture mainly includes three layers, namely the terminal layer, the edge layer, and the DT layer. In the terminal layer, each terminal executes the local model training process and uploads the model parameters to the edge layer. The edge layer includes a base station (BaseStation, BS) and an edge server deployed at the same location as the base station to provide access and computing services. Based on the local model parameters collected from the terminals, the edge server executes the edge aggregation process and obtains a global model. At the same time, abnormal model detection (Abnormal Model Recognition, AMR) is performed on the edge server to identify abnormal models uploaded by malicious terminals to achieve security awareness. The DT layer is constructed on the edge server and maintains synchronization with the physical network through real-time interaction with the terminals to assist the edge server in optimizing resource scheduling.

[0116] As Figure 1As shown, u1, u2, u4, and u6 are scheduled by the base station for local training, and the trained model parameters and status information are sent to the edge server for edge aggregation. The edge server detects that there is a transmission error in the model parameters uploaded by u1, so u1 is rejected from participating in edge aggregation. During edge aggregation, the edge server executes AMR and detects that u6 is a malicious terminal. Therefore, u6 will be blocked from participating in the local training and model parameter upload process.

[0117] (1) Local training model

[0118] Consider a scenario that includes I power Internet of Things terminals, and the set is represented as Y = {u1, u2,..., u I}. The total time period is divided into T iterations, and the set is represented as Τ = {1, 2,..., T}. Assume that the Channel State Information (CSI) remains unchanged within a single iteration but changes between iterations. At the t-th iteration, u i downloads the global model of the (t - 1)-th iteration as the local model M i , and sets the local model parameters as the global model parameters, that is, ω i (t - 1) = ω g (t - 1). Δ i is the dataset for terminal u i to train the local model M i . Define ξ i as the number of CPU cycles required for u i to train a single sample of data, and f i (t) represents the computing resources allocated for local training by u i . Therefore, the energy consumption and delay of local computing can be calculated as

[0119]

[0120] where α is the energy consumption coefficient, and D i represents the size of the local dataset Δ i of u i .

[0121] Define and to represent the input and target output of a sample in the dataset Δ i . The single-sample loss function is used to characterize the deviation between the estimated output and the target output. Therefore, the loss function of u i can be calculated as

[0122]

[0123] The loss function can reflect the accuracy of the local model and be used to guide the update of the parameter ω i (t), that is

[0124]

[0125] where is the learning rate.

[0126] (2) Transmission model

[0127] It is assumed that the base station can select N(t) << I terminals for model upload in each iteration. Define a terminal as an unavailable terminal when its energy is exhausted or it is identified as a malicious terminal by the base station. Define the set of available terminals in the t-th iteration as Σ(t) ∈ Y. Define the terminal access scheduling binary indicator variable as x i (t) ∈ {0, 1}, x i (t) = 1 indicates that the terminal u i is selected for model parameter upload in iteration t, otherwise x i (t) = 0. When x i (t) = 1, u i the local model parameter ω i (t) and the state s i are uploaded to the base station to construct its DT, that is

[0128] DT i = Γ(M i , Δ i , s i ) (4)

[0129] Define |ω i (t)| to represent the magnitude of the parameter ω i (t). Since s i is very small, its upload delay is ignored. Let B i represent the sub-channel bandwidth, P i (t) represent the transmission power, and the transmission rate can be expressed as

[0130]

[0131] where h i (t) represents the channel gain, and N0 and I i (t) represent the noise power and electromagnetic interference power respectively. Therefore, the transmission energy consumption and delay can be expressed as

[0132]

[0133] (3) Packet error rate model

[0134] The edge server uses Cyclic Redundancy Check (CRC) to detect whether the received packet has an error. The packet error rate can be expressed as

[0135]

[0136] where k is a constant related to the coding gain. Define the binary indicator variable a i (t) ∈ {0, 1}, a i (t) = 1 indicates that there is no packet error in the transmission of terminal u i , otherwise a i (t) = 0. a i (t) can be calculated as

[0137]

[0138] (4) Edge aggregation and AMR model

[0139] When all the selected terminals complete the upload of model parameters, the edge server executes the edge aggregation and AMR processes.

[0140] 1) Edge aggregation: Define and to represent the computing resources allocated by the edge server for edge aggregation and AMR respectively. The global model aggregation formula is:

[0141]

[0142] The edge aggregation delay is calculated as

[0143]

[0144] where C0 represents the number of CPU cycles required for the edge server to process 1 bit of data. The loss function of the global model is expressed as

[0145]

[0146] where represents the total number of samples of all selected terminals.

[0147] 2) AMR: To prevent malicious terminals from uploading false model parameters, the present invention uses AMR to detect the accuracy of the model uploaded by the terminal. Define Δ test as the dataset used by the edge server to detect the model, and the loss function of the model parameters ω i of u i (t) on Δ test can be expressed as

[0148]

[0149] Global model ω g (t - 1) at Δ test The loss function can be expressed as

[0150]

[0151] Define when F i,test (ω i (t), t) - F g,test (ω g (t - 1), t) > ι, then u i The uploaded model parameter ω i (t) is abnormal, that is, Z i (t) = 1. Where Z i (t) ∈ {0, 1} is the abnormal model binary indicator variable. The AMR delay can be calculated as

[0152]

[0153] Among them, ξ0 represents the number of CPU cycles required for the edge server to process a single sample. When u i The proportion of the uploaded abnormal models is greater than υ, that is

[0154]

[0155] Determine u i as a malicious terminal and remove it from the set Σ(t + 1).

[0156] (5) Energy consumption and delay model

[0157] u i The total energy consumption is the sum of the local computing energy consumption and the transmission energy consumption, calculated as

[0158]

[0159] The total delay of a single iteration includes the maximum delay of the sum of the local computing delay and the parameter transmission delay among all terminals, the larger delay of the edge aggregation delay and the AMR delay, and the global model broadcast delay τ b , expressed as

[0160]

[0161] Among them,

[0162] 2. Problem modeling and transformation

[0163] (1) Access priority constraint

[0164] According to the business importance and service quality requirements, the terminals have different access priorities. The terminals with high access priorities need to complete a sufficient number of local training and parameter transmissions to ensure the consistency between their DT and the local model. Therefore, the long-term access priority constraint is defined as

[0165]

[0166] where e i is the minimum proportion requirement for selecting u i to perform local training and parameter transmission times.

[0167] (2) Energy consumption constraint

[0168] Considering the limited battery capacity of the terminals, the long-term energy consumption constraint is defined as

[0169]

[0170] where E i,max is the energy budget of u i .

[0171] (3) Problem modeling

[0172] The optimization objective is defined as minimizing the weighted sum of the DT loss function and the total iterative delay by jointly optimizing the computing resource allocation, power control, and access scheduling under the long-term constraints of access priority and energy consumption, expressed as

[0173]

[0174] where V is a non-negative weight, f = (f i (t): u i ∈Σ(t), t∈Τ) is the terminal computing resource allocation vector, P = (P i (t): u i ∈Σ(t), t∈Τ) is the transmission power vector, x = (x i (t): u i ∈Σ(t), t∈Τ) is the terminal access scheduling vector, is the server computing resource allocation vector. C1 and C2 indicate that the BS can schedule at most N(t) terminals to upload model parameters simultaneously. C3 represents the terminal computing resource allocation constraint, where f i,max (t) is the maximum available computing resource of u i . C4 is the transmission power constraint, where P i,max is the maximum transmission power of u i . C5 and C6 are the server computing resource allocation constraints, where is the maximum available computing resource of the edge server. C7 and C8 are the long-term constraints of access priority and energy consumption respectively.

[0175] Based on the relationship between the packet error rate and the DT loss function, P1 can be rewritten as

[0176]

[0177] where

[0178] (4) Problem transformation

[0179] The concept of virtual queues is adopted to decouple the long-term constraints and the short-term resource scheduling optimization. C7 and C8 can be transformed into queue stability constraints, expressed as

[0180] F i (t + 1) = max{F i (t) + e i -x i (t), 0}, (19)

[0181]

[0182] If N i (t) and Y i (t) are average rate stable, then C7 and C8 are automatically satisfied. P2 can be transformed into

[0183]

[0184] s.t. C1~C6,

[0185] C9: F i (t) and Y i (t) are average rate stable. (21)

[0186] Define the vector Θ(t) = [F i (t), Y i (t)], and the Lyapunov function can be expressed as

[0187]

[0188] The Lyapunov drift is defined as the conditional expectation change of the Lyapunov function over two consecutive time slots, expressed as

[0189] ΔL(Θ(t)) = E[L(Θ(t + 1)) - L(Θ(t))|Θ(t)] (23)

[0190] To achieve the minimization of the weighted sum of the global loss function and the total iterative delay under the queue stability constraints, the drift-plus-penalty is defined as

[0191] Δ VL(Θ(t)) = ΔL(Θ(t)) + VE[Φ(t)|Θ(t)] (24)

[0192] where V is the weight of the P3 optimization objective. Therefore, P3 can be transformed into the problem of minimizing the upper bound of Δ V L(Θ(t)) under the constraints C1 - C6, and the upper bound of Δ V L(Θ(t)) is expressed as

[0193]

[0194] 3. Secure and Latency-Aware Digital Twin-Assisted Resource Scheduling Algorithm

[0195] The present invention proposes a secure and latency-aware digital twin-assisted resource scheduling (SAINT) algorithm to solve the joint optimization problem P3, which can be decomposed into two sub-problems, namely SP1: the joint optimization sub-problem of terminal access scheduling, terminal computing resource allocation, and power control; SP2: the server computing resource allocation sub-problem.

[0196] (1) Joint Optimization Sub-Problem of Terminal Access Scheduling, Terminal Computing Resource Allocation, and Power Control

[0197] SP1 is expressed as

[0198]

[0199] s.t. C1 - C4. (23)

[0200] SP1 can be solved in two stages, namely access scheduling and joint optimization of computing resources and power control. First, the access scheduling problem can be modeled as a Markov process and solved by digital twin-assisted deep Q-learning (DQN). The state can be obtained through the digital twin, the state space is defined as G(t) = {F(t), Y(t)}, the action space is defined as A(t) = {x1(t), x2(t), …, x I (t)}, and the cost function is defined as the optimization objective of SP1. DQN can extract features from historical data using a deep neural network and learn the optimal policy based on the estimated Q value Q(G(t), ν(t)), where ν(t) is the set of deep neural network parameters.

[0201] Based on the first stage, the joint optimization problem of computing resources and power control can be defined as SP1'. Since f i,max (t), h i (t) and I i (t) are time-varying and cannot be obtained by the edge server, the present invention uses the empirical values estimated by the digital twin to solve, that is, f′ i,max (t), h′i (t) and I′ i (t). Therefore, SP1' can be expressed as

[0202]

[0203] where

[0204]

[0205] represents the set of scheduled terminals. Based on define SP1' can be approximated as

[0206]

[0207] s.t. C3' and C4. (23)

[0208] SP1” is a convex optimization problem and can be solved by the Lagrangian dual decomposition method.

[0209] (2) Server computing resource allocation sub - problem

[0210] SP2 is expressed as

[0211]

[0212] s.t. C5, C6. (23)

[0213] Since τ a (t) and τ d (t) are inversely proportional to and respectively, and Therefore,[[]]

[0214] When τ a (t) = τ d (t), SP2 reaches the optimum. The optimal server computing resource allocation can be expressed as

[0215]

[0216]

[0217] (3) SAINT

[0218] The algorithm SAINT proposed in the present invention mainly includes access scheduling, model downloading, local training, local model parameter uploading, and AMR and edge aggregation. First, initialize the virtual queue backlog and access scheduling indicator variables to 0. At the beginning of each iteration, the server selects terminals for data transmission based on the estimated Q - value. Secondly, based on the empirical value f′ estimated by digital twin i,max(t), h′ i (t) and I′ i (t), the server optimizes the computing resource allocation and power control by solving SP1”. Subsequently, the selected terminals download the global model ω g (t - 1), and perform local training, parameter transmission, and update F i (t) and Y i (t). Then, by solving SP2 to obtain the optimal computing resource allocation strategy of the server, the server performs edge aggregation and abnormal model identification processes, calculates the computing cost θ(t), and transfers the DQN to the next state G(t + 1). Finally, the server updates the DQN network parameter ν(t) based on the gradient descent method. The framework of SAINT is as Figure 2 shown.

[0219] The present invention conducts simulation experiments on the above-mentioned FEDERATION algorithm, and sets two baseline algorithms for performance comparison and verification. The baseline algorithms are set as follows:

[0220] CS-UCB: An access scheduling algorithm based on the upper confidence bound, which minimizes the total iterative delay through terminal access scheduling optimization. It does not consider energy consumption constraints, computing resource allocation, and power control. All available energy and computing resources are used for parameter transmission and local training.

[0221] RS-DNN: A user association and resource scheduling algorithm based on a deep neural network. This algorithm does not consider access priority constraints, and minimizes the total iterative delay by jointly optimizing terminal computing resource allocation and power control.

[0222] In addition, neither CS-UCB nor RS-DNN considers AMR, and all available resources of the server are used for edge aggregation.

[0223] The simulation results are as Figures 3 to 5 shown:

[0224] Figure 3 shows the change of the DT loss function with the number of iterations. When t = 100, compared with CS-UCB and RS-DNN, SAINT reduces the loss function by 58.06% and 70.39% respectively. The reason is that SAINT uses power control and AMR to mitigate the adverse effects of packet errors and abnormal models on minimizing the DT loss function.

[0225] Figure 4 shows the change of the cumulative iterative delay with the requirement of the DT loss function. When the DT loss function approaches 0.4, compared with CS-UCB and RS-DNN, SAINT reduces the cumulative iterative delay by 16.13% and 38.54% respectively. The reason is that due to considering AMR, SAINT can converge faster.

[0226] Figure 5 It shows the change of the cumulative energy consumption with the number of iterations. When t = 100, compared with CS-UCB and RS-DNN, SAINT can reduce the network cumulative energy consumption by 30.77% and 8.23% respectively. Due to the lack of consideration of the terminal computing resource allocation and power control, CS-UCB has the worst performance.

[0227] The above are only the preferred embodiments of the present invention and are not used to limit the present invention. It should be noted that for those of ordinary skill in the art, without departing from the technical principle of the present invention, several improvements and modifications can be made, and these improvements and modifications should also be regarded as the protection scope of the present invention.

Claims

1. A digital twin-assisted 5G distribution network resource scheduling method, characterized in that The specific scheduling steps are as follows: S1. Build a system model, specifically including: 1.

1. Local training model; 1.

2. Transmit the model; 1.

3. Packet error rate model; 1.

4. Edge aggregation and anomaly model detection model; 1.

5. Energy consumption and delay model; S2. Problem modeling and transformation, specifically including: 2.

1. Access priority constraint; 2.

2. Energy consumption constraint; 2.

3. Problem modeling; 2.

4. Problem transformation; S3. Secure and delay-aware digital twin-assisted resource scheduling method, specifically including: 3.

1. Joint optimization sub-problem of terminal access scheduling, terminal computing resource allocation, and power control; 3.

2. Server computing resource allocation sub-problem; In the access priority constraint in the problem modeling and transformation in S2, the long-term access priority constraint is defined as: where e i is the minimum proportion requirement for selecting u i for local training and the number of parameter transmissions; In the energy consumption constraint in S2, considering the limited terminal battery capacity, the long-term energy consumption constraint is defined as Among them, E i,max is the energy budget of u i ; The optimization objective in the problem modeling in S2 is defined as minimizing the weighted sum of the DT loss function and the total iterative delay through joint optimization of computing resource allocation, power control, and access scheduling under the long-term constraints of access priority and energy consumption, expressed as: C1 and C2 indicate that the BS can schedule at most N(t) terminals to upload model parameters simultaneously; C3 represents the terminal computing resource allocation constraint, where f i,max (t) is the maximum available computing resource of u i ; C4 is the transmission power constraint, where P i,max is the maximum transmission power of u i ; C5 and C6 are the server computing resource allocation constraints, where is the maximum available computing resource of the edge server; C7 and C8 are the access priority and long-term energy consumption constraints respectively; Based on the relationship between the packet error rate and the DT loss function, P1 can be rewritten as: Among them, In the problem transformation in S2, the concept of virtual queues is used to decouple the long-term constraints and short-term resource scheduling optimization. C7 and C8 can be transformed into queue stability constraints, expressed as: F i (t + 1) = max{F i (t) + e i -x i (t), 0}, If N i (t) and Y i (t) are stable in average rate, then C7 and C8 are automatically satisfied, and P2 can be transformed into: Define the vector Θ(t) = [F i (t), Y i (t)], and the Lyapunov function can be expressed as: The Lyapunov drift is defined as the conditional expectation change of the Lyapunov function in two consecutive time slots, expressed as: To minimize the weighted sum of the global loss function and the total iterative delay under the queue stability constraint, the drift-plus-penalty is defined as: Among them, V is the weight of the P3 optimization objective. Therefore, P3 can be transformed into minimizing Δ under the constraints of C1 to C6. V The upper bound problem of L(Θ(t)), Δ V The upper bound of L(Θ(t)) is expressed as:

2. A 5G distribution network resource scheduling method assisted by digital twins according to claim 1, characterized in that: The system model construction in S1 is based on the 5G edge computing distribution network DT architecture of federated learning. The architecture mainly includes three layers, namely the terminal layer, the edge layer, and the DT layer.

3. A 5G distribution network resource scheduling method assisted by digital twin according to claim 1, characterized in that: In the local training model scenario of S1, there are I power Internet of Things terminals, which are represented as a set The total time period is divided into T iterations, which are represented as a set The energy consumption and delay of local computing can be calculated as: Among them, α is the energy consumption coefficient, D i represents u i the size of the local dataset ; f i (t) represents the computing resources allocated for local training of u i ; Definition and represent the input and target output of a sample in the data set, and the single-sample loss function is used to characterize the deviation between the estimated output and the target output. Therefore, the loss function of u can be calculated as follows: i The loss function of can be calculated as: The loss function can reflect the accuracy of the local model and be used to guide the update of the parameter ω i (t), that is: Among them, η is the learning rate.

4. A 5G distribution network resource scheduling method assisted by digital twin according to claim 1, characterized in that: The transmission rate of the transmission model in S1 can be expressed as: B i represents the sub-channel bandwidth, P i (t) represents the transmission power; h i (t) represents the channel gain, and N0 and I i (t) represent the noise power and the electromagnetic interference power respectively; The transmission energy consumption and delay can be expressed as:

5. A 5G distribution network resource scheduling method assisted by digital twin according to claim 1, characterized in that: The packet error rate of the packet error rate model in S1 can be expressed as; where k is a constant related to the coding gain, and a binary indicator variable a i (t) ∈ {0, 1}, a i (t) = 1 indicates that there is no packet error in the transmission of terminal u i Otherwise, a i (t) = 0, a i (t) can be calculated as:

6. The digital twin-assisted 5G distribution network resource scheduling method according to claim 1, wherein: In the edge aggregation in the edge aggregation and anomaly model detection model in S1: The global model aggregation formula is: Define the terminal access scheduling binary indicator variable as x i (t) ∈ {0, 1}; The edge aggregation delay is calculated as: The loss function of the global model is expressed as: Among them represents the total sample quantity of all selected terminals; In the anomaly model detection model in the edge aggregation and anomaly model detection model in S1: The anomaly model detection model delay can be calculated as: where ξ0 represents the number of CPU cycles required for the edge server to process a single sample. When u i the proportion of abnormal models uploaded is greater than υ, that is: Determine u i as a malicious terminal and remove it from the set ; Definition and respectively represent the computing resources allocated by the edge server for edge aggregation and AMR; C0 represents the number of CPU cycles required for the edge server to process 1 bit of data; Definition is the dataset used for the detection model on the edge server side.

7. A 5G distribution network resource scheduling method assisted by digital twin according to claim 1, characterized in that: In the energy consumption and delay model in S1, u i The total energy consumption is the sum of the local computing energy consumption and the transmission energy consumption, which is calculated as: The total delay of a single iteration includes the maximum delay among the sum of the local computing delay and the parameter transmission delay in all terminals, the larger delay between the edge aggregation delay and the AMR delay, and the global model broadcast delay τ b , which is expressed as: Among them, 8. A 5G distribution network resource scheduling method assisted by digital twin according to claim 1, characterized in that: The secure and delay-aware digital twin-assisted resource scheduling algorithm in S3 solves the joint optimization problem P3, which can be decomposed into two sub-problems, namely SP1: Joint optimization sub-problem of terminal access scheduling, terminal computing resource allocation, and power control; SP2: Server computing resource allocation sub-problem; Among them, SP1 is expressed as: SP2 is expressed as: Among them, ν(t) is the set of deep neural network parameters; Since τ a (t) and τ d (t) are inversely proportional to and respectively, and Therefore, When τ a (t) = τ d (t), SP2 reaches the optimum, and the optimal computing resource allocation of the server can be expressed as:

Images