Digital twinning assisted virtual network function resource prediction and migration method

Through the digital twin-assisted virtual network function resource prediction and migration method, combined with the AMT-DMG spatiotemporal fusion prediction model and multi-agent near-end strategy optimization algorithm, the network performance and resource utilization efficiency problems caused by VNF migration are solved, and the physical Maximize network system QoS and migration policy timeliness under resource and latency limitations.

CN119012236AActive Publication Date: 2024-11-22ZHEJIANG NEW INTERNET EXCHANGE CENT CO LTD

Patent Information

Application Number
CN202411091404.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-09
Publication Date
2024-11-22
Estimated Expiration
2044-08-09

AI Technical Summary

Technical Problem

When existing VNF migration methods deal with dynamic changes in network traffic, they lead to loss of status information, affecting network performance and service quality. Traditional methods are difficult to process large-scale data, resulting in delay in migration strategy and limiting the utilization efficiency and performance of network resources.

Method used

Using digital twin-assisted virtual network functional resource prediction and migration method, by constructing a system model including physical infrastructure layer, virtual network layer and digital twin network layer, combining AMT-DMG spatiotemporal fusion prediction model and multi-agent near-end strategy optimization Algorithm predicts VNF resource requirements and formulates migration strategies to ensure the timeliness of migration strategies and stable network performance.

Benefits of technology

Under the limitations of physical resources and end-to-end latency, the QoS of the network system is maximized, network performance and resource utilization efficiency are improved, migration costs and network load balancing problems are reduced, and migration strategies are ensured in time.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119012236A_ABST
    Figure CN119012236A_ABST
Patent Text Reader

Abstract

The invention relates to a digital twinning assisted virtual network function resource prediction and migration method, and belongs to the field of mobile communication. The method comprises the steps that an AMT-DMG space-time fusion prediction model is established to predict the resource demand of a VFN, space-time features are extracted through an attention-based multi-scale time convolution network TCN and a distance-based multi-head map attention network GAT, and fusion is performed to obtain a resource demand prediction result of the VNF; in the training process, a group learning framework based on credibility weighting is embedded into the prediction model, and convergence of the training model is accelerated by sharing training parameters of the prediction model at each VDT node; and according to a prediction result of the VNF resource demand, a multi-agent near-end strategy optimization algorithm is adopted to solve a migration strategy taking maximization of system QoS as an optimization target. According to the invention, the migration probability of the VNF can be reduced under the constraint of guaranteeing the QOS requirement, and the cost required by migration and the network energy consumption are effectively reduced.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of mobile communications and relates to a digital twin-assisted virtual network function resource prediction and migration method. Background Art

[0002] With the development of 5G networks, network slicing technology uses network function virtualization (NFV) and software defined networking (SDN) technologies to divide a single physical network into multiple independent virtual networks, so as to flexibly allocate and manage resources to meet the needs of different application scenarios. NFV virtualizes network functions into virtual network functions (VNFs), dynamically deploys them on different physical nodes, and forms service function chains (SFCs) to provide services. SDN separates the control plane and data plane of the network, allowing the controller to centrally control and dynamically adjust resources. However, the dynamic changes in network traffic have caused VNF migration problems. Especially when physical nodes or links are overloaded, re-instantiating VNFs will cause state information loss, threatening network performance and service quality.

[0003] Existing VNF migration methods mainly rely on historical resource usage data for prediction and pre-planning migration. However, these methods have shortcomings, such as link overload and poor VNF supervision. In addition, most methods make predictions centrally on the SDN controller, which is difficult to process large-scale data, and may cause delays when synchronizing VNF resource demand data, resulting in the inability to adjust the migration strategy in time, affecting the quality of network services. Faced with complex high-dimensional migration problems, traditional heuristic algorithms often cannot find the optimal migration solution, limiting the utilization efficiency and performance of network resources. Summary of the invention

[0004] In view of this, the purpose of the present invention is to provide a digital twin-assisted virtual network function resource prediction and migration method to solve the network energy consumption, migration cost and network load balancing problems brought about by VNF migration, and maximize the QoS of the network system under the constraints of physical resources and end-to-end delay. At the same time, by predicting the future resource requirements of VNF, the migration strategy is formulated in advance to ensure the timeliness of the migration strategy and the stability of network performance.

[0005] In order to achieve the above object, the present invention provides the following technical solutions:

[0006] A digital twin-assisted virtual network function resource prediction and migration method comprises the following steps:

[0007] S1. Build a digital twin-assisted VNF resource prediction and migration system model, which includes the physical infrastructure layer, virtual network layer, and digital twin network layer;

[0008] S2. Establish an end-to-end delay model, system energy consumption model, cost model, load balancing model, information age violation probability model, and the optimization goal of maximizing system QoS based on the VNF resource prediction and migration system model;

[0009] S3. Establish an AMT-DMG spatiotemporal fusion prediction model to predict the resource requirements of VFN, wherein the spatiotemporal features are extracted and fused by the attention-based multi-scale temporal convolutional network TCN and the distance-based multi-head graph attention network GAT to obtain the resource demand prediction results of VNF; during the training process, the credibility-weighted group learning framework is embedded in the AMT-DMG spatiotemporal fusion prediction model, and the convergence of the training model is accelerated by sharing the training parameters of the AMT-DMG spatiotemporal fusion prediction model at each VDT node;

[0010] S4. Based on the prediction results of VNF resource requirements, a multi-agent proximal strategy optimization algorithm is used to solve the migration strategy with maximizing system QoS as the optimization goal.

[0011] Further, in step S1, the physical infrastructure layer provides physical network resources for the virtual network layer, and the physical infrastructure layer is modeled as a connected undirected graph in is the underlying physical node set, is the underlying physical link set, and the CPU resources and storage resources of physical node n are and represents the physical link between physical nodes n and m, and physical link l nm The bandwidth resource is The physical nodes include twin service nodes for building VNF virtual service nodes and twin service nodes for building VNF digital twins.

[0012] The virtual network layer is an undirected graph structure network consisting of virtual nodes mapped by multiple physical devices and virtual links mapped by physical links. It builds a service function chain SFC according to the business requests of different users. Suppose a set of SFCs is represented as Each SFC is a set of VNFs arranged in order, where represents the i-th SFC, represents the VNF set of the i-th SFC, represents the jth VNF on the i-th SFC, for The virtual link set is Abstract as a directed graph The CPU resource requirements and storage resource requirements are and Virtual Link express and Virtual link between The bandwidth resource requirement is Defining binary variables Used to reflect the mapping relationship between virtual network and physical network. When At time t, it is mapped to the physical node n. When At time t, it is not mapped to physical node n; when When , it indicates a virtual link Mapped to physical link l at time t nm On, when When , it indicates a virtual link At time t, there is no mapping to the physical link l nm superior.

[0013] The digital twin network layer builds a digital twin network of the virtual network by collecting valid status information in the physical network and the virtual network. Each VNF node will be associated with a digital twin service node and build a corresponding digital twin. The digital twin of the VNF is called a VDT node. The future resource demand results of the VDT node are predicted based on the historical resource data in the data collection library to determine whether the VNF node will have an abnormal situation at the next moment, formulate a VNF migration strategy in advance, and then send the strategy to the physical network through the SDN controller for implementation.

[0014] Further, in step S2, for the end-to-end delay model, an SFC service request The end-to-end delay τ at time t total (t) is expressed as:

[0015]

[0016] in, yes size; is the number of network routing hops between physical node m and physical node n; t is the delay required to transmit one unit of data along the migration path, that is, the transmission rate of VNF migration; Θ is a positive coefficient; Indicates physical link l nm The physical distance; c is the speed of light; express The processing rate; is the processing rate coefficient of the data packet; Service Request for SFC The packet size.

[0017] For the system energy consumption model, the total energy consumption of the network at time t is P total (t) is expressed as:

[0018]

[0019] Among them, the energy consumption of the underlying physical node n includes the energy consumption in the active state Energy consumption in sleep mode and the energy consumption of switching states B n (t) is the switch binary variable of the physical node, B n (t) = 1 means the node is turned on, Z n (t) is a binary variable indicating whether the state of the physical node n has changed at time t, expressed as:

[0020]

[0021] Among them, Z n (t) = 1 means that the state of the management node n changes at time t, Z n (t)=0 indicates that the state of the management node n at time t has not changed.

[0022] For the cost model, when the VNF instance carried by a physical node needs to be migrated to another physical node that can meet the VNF resource requirements, at time t, the overall network migration cost is expressed as:

[0023]

[0024] Among them, κ is the migration cost factor, Is a node The size of is the number of network routing hops between physical node m and physical node n.

[0025] For the load balancing model, the variance of the CPU resource load, storage resource load, and link bandwidth resource load in the network can be expressed as:

[0026]

[0027]

[0028]

[0029] in, and Indicates the CPU resource and storage resource load of physical node n at time t, For physical link bandwidth resource load, and It represents the average value of the system's CPU resource load, storage resource load, and link bandwidth resource load at time t, and the load balancing capability of the network system L total (t) is expressed as:

[0030]

[0031] Among them, μ1=μ2=μ3=1 / 3, that is, CPU resources, storage resources and bandwidth resources have the same impact on network load balancing;

[0032] For the information age violation probability model, assume that the generation time of the Mth data packet of the VDT synchronization information is t M , the time when the VDT node receives the update information packet is t' M , X M =t M -t M-1 is the time interval between two adjacent data packets, T M =t' M -t M is the time the data packet spends in the system, Y M =t' M -t' M-1 is the time interval between two adjacent data packets arriving at the VDT node, is the AoI constraint threshold, then at time t, the instantaneous information age Δ(t), t>0 is expressed as:

[0033] Δ(t)=t-max{t M :t' M ≤t}

[0034] Among them, max{t M :t' M ≤t} is the timestamp of the most recent update information received by the VDT node before time t; G(M) indicates that the AoI of the Mth synchronization data packet is greater than the threshold constraint The duration of G(M) is expressed as:

[0035]

[0036] The AoI violation probability is defined as the AoI of the VDT node successfully receiving the update information and the AoI exceeding the given AoI threshold. The probability is expressed as:

[0037]

[0038] in, The number of times the VDT node receives update information packets.

[0039] Further, in step S2, the QoS of the system is expressed as:

[0040] Q=ξ1Q Cost +ξ2Q Power +ξ3Q Load

[0041] Among them, Q Cost , Q Power and Q Load They represent the QoS of system migration cost, the QoS of system energy consumption, and the QoS of system load balancing, respectively. ξ1, ξ2, and ξ3 are the weight factors of migration cost, energy consumption, and load balancing indicators on system QoS. ξ1+ξ2+ξ3=1 and satisfy ξ1,ξ2,ξ3∈(0,1). The optimization objective is subject to the following constraints:

[0042] C1:

[0043] C2:

[0044] C3:

[0045] C4:τ total (t)≤τ req

[0046] C5:

[0047] C6:

[0048] C7:

[0049] C8:

[0050] C9:

[0051] C10:

[0052] C11:

[0053] C12:

[0054] C13:

[0055] C14:

[0056] Among them, C1 ensures that each VNF is mapped to only one physical node; C2 ensures that each virtual link is mapped to only one physical link; C3 represents the AoI violation probability limit of the twin synchronization data, Pr max is the maximum AoI violation probability; C4 represents the end-to-end delay limit of SFC, Request for SFC The end-to-end delay requirement is 100%. C5 indicates that the physical node carrying the VNF is in the turned-on state. C6-C8 indicate that the CPU resources and storage resources of the VNF cannot exceed the resource limits of the physical node, and the virtual link cannot exceed the bandwidth resource limit of the physical link. and are the upper and lower limit factors of resource utilization of physical network resources; C9-C10 indicate that when a virtual link is mapped to a physical link, the corresponding VNF node must also be mapped to the corresponding physical node; C11-14 are binary variable constraints during VNF migration.

[0057] Further, in step S3, the input time series data is x={x(t)|t=1,2,...,T′}, T′ is the total length of the time step of the data, and the feature data is x(t)={x j (t)|j=1,2,...,N′}, N′ is the total number of VDT nodes, x j (t) is the characteristic data of the jth VNF, and finally the characteristic data of the VNF at time t+1 is output x(t+1)={x j (t+1)|j=1,2,...,N′};

[0058] Among them, the process of extracting temporal features by the attention-based multi-scale temporal convolutional network TCN model is:

[0059] The TCN model includes TCN residual blocks with filters of different sizes, each TCN residual block includes two dilated causal convolutions and weight normalization, and outputs the results through residual connections and ReLU activation functions, where the calculation formula for the dilated causal convolution is:

[0060]

[0061] In the formula, x(t) is the input time series data feature, For the The output of the dilated causal convolution in the TCN, For the The convolution kernel size of a TCN, f(g) is the g-th element in the convolution kernel, x(td g) represents the data at different time steps, and d is the dilation coefficient;

[0062] The output of the residual connection is the temporal features extracted by the TCN network at different time scales. The calculation formula is as follows:

[0063]

[0064] The calculation process of the time scale attention layer is as follows:

[0065]

[0066]

[0067] In the formula, w1 and w2 are weight coefficients, b is the bias coefficient, To normalize the attention weight value, the output of the attention layer is used express;

[0068] The results of the TCN networks at different time scales after importance weighting are spliced, and then input into the average pooling layer and the fully connected layer to obtain the time feature matrix H t :

[0069]

[0070] Among them, concat(·) represents the concatenation function, Avgpooling(·) is the average pooling layer operation, and FC(·) is the fully connected layer operation.

[0071] Further, in step S3, the distance-based multi-head graph attention network GAT model extracts spatial features as follows:

[0072] Define the node feature input vector as:

[0073]

[0074] in, is the feature representation of node u, is the feature vector dimension of the node, U is the number of VNF nodes in the network;

[0075] The state feature update of each node is expressed as:

[0076]

[0077] Among them, δ(*) is a nonlinear activation function, and the feature representation of the central node u is updated through aggregation operation as follows: Λ u Represents the label set of neighbor nodes of node u, Φ' uu′ is the attention coefficient, expressed as:

[0078]

[0079] Among them, ε is the attention coefficient threshold, d' uu′ is the reciprocal normalized distance, Φ uu′ is the normalized attention coefficient of the influence of neighbor node u′ on the central node u;

[0080] Introduce a multi-head attention mechanism to focus on multiple different features of the input. Suppose the number of attention heads of the attention mechanism is K head , then the updated feature representation of the central node is:

[0081]

[0082] Aggregate the update results of each node to obtain the spatial feature matrix of the VNF node

[0083] The time feature matrix H t With the spatial feature matrix H s The input is integrated into the fully connected layer for fusion and dimension reduction to obtain the spatiotemporal feature matrix representing the VNF resource demand characteristics, and then the prediction result of the VNF resource demand at the next moment is obtained:

[0084] H=FC2(FC1(H t ,H s ))

[0085] H={x j (t+1)|j=1,2,...,N′}

[0086] Where H = FC2(FC1(H t ,H s )) is the fused spatiotemporal feature matrix, FC1 and FC2 are different fully connected layers; H = {x j (t+1)|j=1,2,...,N′} is the prediction result of VNF resource demand at the next moment.

[0087] Further, in step S3, the process of embedding the credibility-weighted group learning framework in the AMT-DMG spatiotemporal fusion prediction model is as follows:

[0088] 1) Each VDT node has a local AMT-DMG spatiotemporal fusion prediction model. First, each VDT node runs its own AMT-D spatiotemporal fusion prediction model locally and uploads its own local model training parameters to the Swarm Learning (SL) network:

[0089] The SL network initializes the global model parameters as θ, and the jth VDT node combines its local sample data set To train local model parameters Use the gradient descent algorithm to update the local training model parameters:

[0090]

[0091] Among them, λ′>0 is the learning rate, is the loss function over parameters The gradient of

[0092] 2) Secondly, the SL network selects a temporary central integration node, which uses a credibility weighted algorithm to merge the model training parameters of all participants:

[0093] The credibility value of each VDT is expressed as is the total number of VDT nodes, and the selection process of the temporary central integration point is expressed as:

[0094]

[0095] After uploading the model parameters of all participants to the temporary central integration node, the model parameters are merged and the global model parameter θ is obtained using the weighted average function. o , expressed as:

[0096]

[0097] 3) Finally, the temporary central integration node distributes the merged model parameters to other participants for training and dynamically updates the local model, repeating the training until the entire training process converges.

[0098] Further, in step S4, the following steps are included:

[0099] S41: Calculate resource utilization of CPU resources, storage resources, and bandwidth resources of physical nodes based on the resource prediction results of the VNF, and select overloaded or underloaded VNF nodes that need to be migrated;

[0100] S42: For VNF migration, the system QoS is optimized to the maximum extent possible under the conditions of meeting the end-to-end delay and the age limit of the digital twin synchronization data information. The migration cost, network energy consumption, and network load balancing are used to measure the system QoS.

[0101] S43: A multi-agent proximal strategy optimization deep reinforcement learning algorithm is used to solve the VNF migration problem. Each SFC is regarded as an agent, and the optimization goal is to maximize the global reward of the system to obtain the optimal VNF migration strategy.

[0102] Further, in step S43, the process of solving the migration strategy by the multi-agent proximal strategy optimization algorithm is as follows:

[0103] Suppose there are K SFC requests, each SFC is defined as an agent, and the MDP model of each agent is defined as Represent the state space Action Space State transition probability Reward Function

[0104] Then the global reward r of the multi-agent system is t o It is expressed as:

[0105]

[0106] Among them, r t k represents the instantaneous reward of agent k;

[0107] Each agent has an Actor-Critic structure, and the Actor network parameters are expressed as The critic network parameters are expressed as They represent the Actor network parameters and Critic network parameters of agent k respectively;

[0108] Introducing a discount factor γ∈[0,1), a global cumulative discount reward Among them, i represents the number of steps from the current time step to the future. represents the global reward at time t+i;

[0109] Let the global state value function be V π (s t ) indicates state s t The long-term total reward under , where π is the strategy, and the state value function is expressed as:

[0110]

[0111] in is the expected operator, then the optimal VNF migration strategy π * It is expressed as:

[0112]

[0113] The Actor network updates the network parameters through the policy loss function. The policy loss function of the Actor network corresponding to agent k is expressed as:

[0114]

[0115] in, is the ratio of the old and new strategies of the Actor network, is an advantage function, clip() is a truncation function, To set the hyperparameters for clipping range of the clip() function;

[0116] The parameter update of the critic network is based on the state value loss function, which is defined as the mean square error between the cumulative discounted reward and the state value function, which is expressed as:

[0117]

[0118] in, is the cumulative discounted reward of agent k, is the state value function of agent k;

[0119] Finally, the gradient ascent method is used to update the parameters of the Actor network, and the gradient descent method is used to update the parameters of the Critic network, which are expressed as:

[0120]

[0121]

[0122] Among them, α a and α c They are the learning rates of the Actor network and the Critic network respectively;

[0123] After the parameters are updated, the Actor network and the Critic network will improve the strategy and state value function according to the new parameters, and continuously iterate the training to optimize the VNF migration strategy.

[0124] Furthermore, for each agent's state space in the MDP model There are: The state space of the kth agent is defined as in represents the network state space of physical nodes, represents the state space of the physical link, and Represents the computing resource requirement set and storage resource requirement set of the VNF in an SFC. is a set of virtual link bandwidth resource requirements;

[0125] For each agent's action space in the MDP model There are: define the action space of the kth agent in A collection of actions mapped to virtual nodes, is a set of virtual link mapping actions;

[0126] For each agent's MDP model, the state transition probability function There are: The state transition probability function represents the agent in state s t Take action a t , and transfer to a new state at the next moment The conditional probability distribution of t ×a t ×s t+1 →[0,1], a set is usually used to represent the conditional probability distribution, which contains the probability of each action taken in the current state to transfer to the next state. The state transition probability is expressed as

[0127] For each agent's MDP model, the reward function At time t, the kth agent is in state Take Action Will generate a strategy According to the optimization objectives and constraints of the VNF migration problem, we can get the instantaneous reward r of agent k: t k for:

[0128]

[0129] Among them, Q Cost , Q Power and Q Load They represent the QoS of system migration cost, the QoS of system energy consumption and the QoS of system load balancing respectively. ξ1, ξ2 and ξ3 are the weight factors of migration cost, energy consumption and load balancing indicators on system QoS. Ω represents the penalty factor, which is an infinitesimal positive real number.

[0130] The beneficial effects of the present invention are:

[0131] In order to solve the problem of frequent VNF ​​migration caused by the time-varying characteristics of network services, the present invention proposes a resource prediction scheme based on the AMT-DMG spatiotemporal fusion feature extraction of a credibility-weighted group learning framework. This scheme improves the accuracy of resource prediction results by fusing time dimension features and space dimension features. At the same time, the group learning framework is used to alleviate the data pressure and memory shortage problems caused by centralized training, improve the training efficiency of the model, and based on the resource demand prediction results of VNF, a multi-agent-based proximal strategy optimization VNF migration algorithm is proposed to optimize the system QoS and network performance under the constraint of ensuring service delay.

[0132] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0133] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below in conjunction with the accompanying drawings, wherein:

[0134] Figure 1 This is the system architecture diagram of VNF resource prediction and migration assisted by digital twins;

[0135] Figure 2 This is the structure diagram of the AMT-DMG spatiotemporal fusion prediction model;

[0136] Figure 3 It is a credibility-weighted group learning framework diagram;

[0137] Figure 4 It is the network structure diagram of the multi-agent proximal strategy optimization algorithm. DETAILED DESCRIPTION

[0138] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0139] Among them, the drawings are only used for illustrative explanations, and they only represent schematic diagrams rather than actual pictures, and should not be understood as limitations on the present invention. In order to better illustrate the embodiments of the present invention, some parts of the drawings may be omitted, enlarged or reduced, and do not represent the size of actual products. For those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.

[0140] The same or similar numbers in the drawings of the embodiments of the present invention correspond to the same or similar parts; in the description of the present invention, it should be understood that if the terms "upper", "lower", "left", "right", "front", "rear", etc. indicate the orientation or position relationship, they are based on the orientation or position relationship shown in the drawings, which is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operate in a specific orientation. Therefore, the terms describing the position relationship in the drawings are only used for illustrative purposes and cannot be understood as limiting the present invention. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances.

[0141] See also Figure 1 to Figure 4 , which is a digital twin-assisted virtual network function resource prediction and migration method.

[0142] Example

[0143] This embodiment proposes a digital twin-assisted virtual network function resource prediction and migration method, which includes the following steps:

[0144] S1. Build a digital twin-assisted VNF resource prediction and migration system model, which includes the physical infrastructure layer, virtual network layer, and digital twin network layer;

[0145] S2. Establish an end-to-end delay model, system energy consumption model, cost model, load balancing model, information age violation probability model, and the optimization goal of maximizing system QoS based on the VNF resource prediction and migration system model;

[0146] S3. Establish an AMT-DMG spatiotemporal fusion prediction model to predict the resource requirements of VFN, wherein the spatiotemporal features are extracted and fused by the attention-based multi-scale temporal convolutional network TCN and the distance-based multi-head graph attention network GAT to obtain the resource demand prediction results of VNF; during the training process, the credibility-weighted group learning framework is embedded in the AMT-DMG spatiotemporal fusion prediction model, and the convergence of the training model is accelerated by sharing the training parameters of the AMT-DMG spatiotemporal fusion prediction model at each VDT node;

[0147] S4. Based on the prediction results of VNF resource requirements, a multi-agent proximal strategy optimization algorithm is used to solve the migration strategy with maximizing system QoS as the optimization goal.

[0148] In this embodiment, the physical infrastructure layer in step S1 provides physical network resources for the virtual network layer, which are mainly divided into CPU resources, storage resources of the underlying physical devices, and bandwidth resources of the physical links. The physical infrastructure layer can be modeled as a connected undirected graph in is the underlying physical node set, is the underlying physical link set, and the CPU resources and storage resources of physical node n are and represents the physical link between physical nodes n and m, and physical link l nm The bandwidth resource is Physical nodes are divided into two types: one is the physical node used for virtual network function mapping, which is the virtual service node used to build VNF; the other is the physical node that builds the VNF digital twin by collecting data information of the physical network and the virtual network, which is the twin service node.

[0149] The virtual network layer is an undirected graph structure network consisting of virtual nodes mapped to multiple physical devices and virtual links mapped to physical links. It builds a service function chain (SFC) according to the business requests of different users. Suppose a set of SFCs is represented as Each SFC consists of a set of VNFs arranged in order. represents the i-th SFC, represents the VNF set of the i-th SFC, for The set of virtual links, then It can be abstracted into a directed graph The CPU resource requirements and storage resource requirements are and Virtual Link express and Virtual link between The bandwidth resource requirement is We define the binary variable Used to reflect the mapping relationship between virtual network and physical network. When At time t, it is mapped to physical node n, and vice versa. When , it indicates a virtual link Mapped to physical link l at time t nm and vice versa.

[0150] The digital twin network layer collects valid status information from the physical network and the virtual network to build a digital twin network of the virtual network. Each VNF node needs to be associated with a digital twin node and build its own twin on it, called a VDT node. The future resource demand results of the VDT node are predicted based on the historical resource data in the data collection library to determine whether the VNF node will be underloaded, overloaded, or faulty at the next moment. VNF migration strategies are formulated in advance and then sent to the physical network through the SDN controller for implementation, so as to achieve the purpose of maintaining the performance of the real physical network.

[0151] End-to-end delay model: an SFC service request The end-to-end delay τ at time t total (t) can be expressed as:

[0152]

[0153] in, yes The size of is the number of network routing hops between physical node m and physical node n, Ψ t is the delay required to transmit a unit of data along the migration path, that is, the transmission rate of VNF migration, Θ is a positive coefficient, Indicates physical link l nm The physical distance, c is the speed of light, express The processing rate, is the processing rate coefficient of the data packet, for The requested packet size.

[0154] The system energy consumption model established in step S2 of this embodiment is:

[0155] The total energy consumption of the network at time t is P total (t), can be expressed as:

[0156]

[0157] Among them, the energy consumption of the underlying physical node n includes the energy consumption in the active state Energy consumption in sleep mode and the energy consumption of switching states B n (t) is the switch binary variable of the physical node, B n (t) = 1 means the node is turned on, Z n (t) is a binary variable indicating whether the state of the physical node n has changed at time t, which can be expressed as:

[0158]

[0159] Among them, Z n (t) = 1 means that the state of the management node n changes at time t, Z n (t)=0 indicates that the state of the management node n at time t has not changed.

[0160] The cost model established in step S2 of this embodiment is:

[0161] When the server is overloaded or underloaded, the VNF instances carried by the physical node need to be migrated to other physical nodes that can meet the VNF resource requirements. At time t, the overall network migration cost can be expressed as:

[0162]

[0163] Among them, κ is the migration cost factor, yes The size of is the number of network routing hops between physical node m and physical node n.

[0164] The load balancing model established in step S2 of this embodiment is:

[0165] The variance of CPU resource load, storage resource load, and link bandwidth resource load in the network can be expressed as:

[0166]

[0167]

[0168]

[0169] in, and Indicates the CPU resource and storage resource load of physical node n at time t, For physical link bandwidth resource load, and It represents the average value of the system's CPU resource load, storage resource load, and link bandwidth resource load at time t, and the load balancing capability of the network system L total (t) can be expressed as:

[0170]

[0171] Among them, μ1=μ2=μ3=1 / 3, that is, CPU resources, storage resources and bandwidth resources have the same impact on network load balancing.

[0172] The Age of Information (AoI) violation probability model established in step S2 of this embodiment is:

[0173] Assume that the generation time of the Mth data packet of VDT synchronization information is t M , the time when the VDT node receives the update information packet is t' M , X M =t M -t M-1 is the time interval between two adjacent data packets, T M =t' M -t M is the time the data packet spends in the system, Y M =t' M -t' M-1 is the time interval between two adjacent data packets arriving at the VDT node, is the AoI constraint threshold, then at time t, the instantaneous AOI can be recorded as Δ(t) (t>0) and expressed as:

[0174] Δ(t)=t-max{t M :t' M ≤t}

[0175] Among them, max{t M :t' M ≤t} is the timestamp of the most recent update information received by the VDT node before time t. G(M) indicates that the AoI of the Mth synchronization data packet is greater than the threshold constraint The duration of G(M) can be expressed as:

[0176]

[0177] The AoI violation probability is defined as the AoI of the VDT node successfully receiving the update information and the AoI exceeding the given AoI threshold. The probability can be expressed as:

[0178]

[0179] in, The number of times the VDT node receives update information packets.

[0180] The QoS of the system established in step S2 of this embodiment is:

[0181] The QoS of the system is related to the migration cost, energy consumption, and load balancing capability, which can be expressed as:

[0182] Q=ξ1Q Cost +ξ2Q Power +ξ3Q Load

[0183] In the above formula, Q Cost , Q Power and Q Load They represent the QoS of system migration cost, the QoS of system energy consumption, and the QoS of system load balancing, respectively. ξ1, ξ2, and ξ3 are the weight factors of migration cost, energy consumption, and load balancing indicators on system QoS. ξ1+ξ2+ξ3=1 and satisfy ξ1,ξ2,ξ3∈(0,1). The optimization objective is subject to the following constraints:

[0184] C1:

[0185] C2:

[0186] C3:

[0187] C4:τ total (t)≤τ req

[0188] C5:

[0189] C6:

[0190] C7:

[0191] C8:

[0192] C9:

[0193] C10:

[0194] C11:

[0195] C12:

[0196] C13:

[0197] C14: Among them, C1 ensures that each VNF is mapped to only one physical node, C2 ensures that each virtual link is mapped to only one physical link, and C3 represents the AoI violation probability limit of the twin synchronization data. max is the maximum AoI violation probability, C4 represents the end-to-end delay limit of SFC, for The end-to-end delay requirement is C5, which means that the physical node carrying the VNF is in the open state. C6-C8 means that the CPU resources and storage resources of the VNF cannot exceed the resource limit of the physical node, and the virtual link cannot exceed the bandwidth resource limit of the physical link. and are the upper and lower limit factors of resource utilization of physical network resources. C9-C10 indicate that when a virtual link is mapped to a physical link, the corresponding VNF node must also be mapped to the corresponding physical node. C11-14 are binary variable constraints during VNF migration.

[0198] In step S3, this embodiment establishes Figure 2 The AMT-DMG spatiotemporal fusion prediction model shown in the figure extracts spatiotemporal features and fuses them through the attention-based multi-scale TCN model and the distance-based multi-head GAT model to obtain the resource demand prediction result of VNF. The input time series data is x = {x(t)|t = 1, 2, ..., T'}, T' is the total length of the time step of the data, and the feature data is x(t) = {x j (t)|j=1,2,...,N′}, N′ is the total number of VDT nodes, x j (t) is the characteristic data of the jth VNF, including the CPU and storage capacity of the node. Through the AMT-DMG spatiotemporal fusion prediction model, the relevant features of the time dimension and spatial dimension of the data are extracted and fused respectively, and finally the characteristic data of the VNF at time t+1 is output x(t+1)={x j (t+1)|j=1,2,...,N′}.

[0199] Specifically, step S3 includes the following steps:

[0200] S31: Collect network history resource information of physical and virtual networks through digital twin networks;

[0201] S32: In the digital twin network, each VDT node has a local AMT-DMG spatiotemporal fusion prediction model, and the prediction model is trained using the collected network historical resource information;

[0202] S33: A multi-scale TCN based on the attention mechanism is used to extract the temporal features of VNF resource demand data. The feature representation capability of the model at multiple time scales is enhanced by splicing TCNs of different scales. At the same time, time scale attention is introduced to weight the importance of TCNs of different scales. By reducing the weights of unimportant time scale features, the interference of these features on the prediction results can be reduced.

[0203] S34: A distance-based multi-head GAT model is used to extract the spatial features of VNF resource demand data. The virtual network is modeled as a graph structure, and the features of neighbor nodes are aggregated to update the feature representation of the central node, thereby updating the feature representation of the entire network. A multi-head attention mechanism is introduced to focus on multiple different features of different neighbor nodes and assign importance weights. At the same time, the influence of the distance between nodes on the attention weight is also considered, so as to achieve the effect of improving the prediction accuracy.

[0204] S35: Embed the credibility-weighted group learning framework into the prediction model. The SL framework reduces the burden of the central integration node by sharing model parameters. Each node can also benefit from the training results of other nodes, promote collaboration and knowledge sharing, accelerate the convergence of the training model, and improve the efficiency of model training. At the same time, considering that each participant in group learning has a different degree of influence on the merged model parameters, credibility is used to measure the differences between different participants, and a temporary central integration point for parameter aggregation is selected according to the credibility, so as to improve the accuracy of the global model.

[0205] The process of extracting temporal features based on the attention-based multi-scale TCN model can be described as follows: Assuming that TCN residual blocks with filters of different sizes, each TCN residual block mainly consists of two dilated causal convolutions and weight normalization, and outputs the results through residual connections and ReLU activation functions.

[0206] The calculation formula of dilated causal convolution is as follows:

[0207]

[0208] In the formula, x(t) is the input time series data feature, For the The output of the dilated causal convolution in the TCN, For the The convolution kernel size of a TCN, f(g) is the g-th element in the convolution kernel, x(td g) represents the data at different time steps, and d is the dilation coefficient.

[0209] The output of the residual connection is the temporal features extracted by the TCN network at different time scales. The calculation formula is as follows:

[0210]

[0211] The calculation process of the time scale attention layer is as follows:

[0212]

[0213]

[0214] In the formula, w1 and w2 are weight coefficients, b is the bias coefficient, To normalize the attention weight value, the output of the attention layer is used The results of the TCN networks at different time scales after importance weighting are concatenated using the concat() function, and then input into the average pooling layer Avgpooling and the fully connected layer FC to obtain the time feature matrix H t :

[0215]

[0216] The process of extracting spatial features based on the distance-based multi-head GAT model can be described as: Defining the node feature input vector can be described as in, is the feature representation of node u, is the feature vector dimension of the node, and U is the number of VNF nodes in the network. The state feature update of each node can be expressed as:

[0217]

[0218] Among them, δ(*) is a nonlinear activation function, and the feature representation of the central node u is updated through aggregation operation as follows: Λ u Represents the label set of neighbor nodes of node u, Φ' uu′ is the attention coefficient, which can be expressed as:

[0219]

[0220] Among them, ε is the attention coefficient threshold, d' uu′ is the reciprocal normalized distance, Φ uu′ It is the normalized attention coefficient of the influence of neighbor node u′ on the central node u.

[0221] In addition, the introduction of the multi-head attention mechanism can focus on multiple different features of the input at the same time, alleviating the overfitting problem. head is the number of heads of the attention mechanism, then the feature representation of the updated central node can be further expressed as:

[0222]

[0223] Aggregate the update results of each node to obtain the spatial feature matrix of the VNF node

[0224] The spatiotemporal feature fusion process can be described as follows: the obtained temporal feature matrix and spatial feature matrix are input into the fully connected layer for fusion and dimension reduction, and the spatiotemporal feature matrix H = FC2(FC1(Ht ,H s )), we get the prediction result of VNF resource demand at the next moment:

[0225] H={x j (t+1)|j=1,2,...,N′}

[0226] Additionally, in step S3, this embodiment embeds a credibility-weighted group learning framework into the AMT-DMG spatiotemporal fusion prediction model, such as Figure 3 In the credibility-weighted group learning framework shown in the figure, each VDT node has a local AMT-DMG model. First, each AMT-DMG model is run locally, and the training parameters of each local model are uploaded to the SL network. Then, the SL network selects a temporary central integration node, and then the temporary central integration node uses a credibility-weighted algorithm to merge the model training parameters of all participants. Finally, the temporary central integration node distributes the merged model parameters to other participants for training and dynamically updates the local model. The training is repeated until the entire training process converges.

[0227] The SL network initializes the global model parameters as θ, and the jth VDT node combines its local sample data set To train local model parameters Use the gradient descent algorithm to update the local training model parameters:

[0228]

[0229] Among them, λ′>0 is the learning rate, is the loss function over parameters gradient.

[0230] The credibility value of each VDT is expressed as is the total number of VDT nodes, the selection process of the temporary central integration point can be expressed as:

[0231]

[0232] After uploading the model parameters of all participants to the temporary central integration node, the model parameters are merged and the global model parameter θ is obtained using the weighted average function. o , which can be expressed as:

[0233]

[0234] In step S4 of this embodiment, a VNF migration strategy is formulated according to the prediction results of the AMT-DMG spatiotemporal fusion prediction model. Step S4 includes the following steps:

[0235] S41: Calculate the resource utilization of the physical node through the resource prediction results of the VNF, including CPU resources, storage resources, and bandwidth resources, so as to screen out the overloaded or underloaded VNF nodes that need to be migrated.

[0236] S42: Through the migration of VNFs, the system QoS is optimized while ensuring the end-to-end delay and the age limit of digital twin synchronization data information. The system QoS is measured by three indicators: migration cost, network energy consumption, and network load balancing.

[0237] S43: Considering the complexity of the VNF migration problem, a deep reinforcement learning algorithm with multi-agent proximal strategy optimization is used to solve the VNF migration problem. Each SFC is regarded as an agent, and the optimization goal is to maximize the global reward of the system to obtain the optimal VNF migration strategy.

[0238] like Figure 4 The figure shows the network structure of the multi-agent proximal strategy optimization algorithm. Assuming there are K SFC requests, this embodiment defines each SFC as an agent, and the MDP model of each agent can be defined as Represent the state space Action Space State transition probability function Reward Function

[0239] (1) State space: The state space of the kth agent is defined as in represents the network state space of physical nodes, represents the state space of the physical link, and Represents the computing resource requirement set and storage resource requirement set of the VNF in an SFC. is the set of virtual link bandwidth resource requirements,

[0240] (2) Action space: Define the action space of the kth agent in A collection of actions mapped to virtual nodes, A collection of virtual link mapping actions.

[0241] (3) State transition probability function: The state transition probability function represents the state of the agent in state s t Take action a t , and transfer to a new state at the next moment The conditional probability distribution of t ×a t ×s t+1→[0,1], a set is usually used to represent the conditional probability distribution, which contains the probability of each action taken in the current state to transfer to the next state. The state transition probability is expressed as

[0242] (4) Reward function: At time t, the kth agent is in state Take Action Will generate a strategy According to the optimization objectives and constraints of the VNF migration problem, the instantaneous reward of agent k can be obtained for:

[0243]

[0244] Among them, Ω is the penalty factor, which is an infinitesimal positive real number.

[0245] The global reward r of the multi-agent system t o It can be expressed as:

[0246]

[0247] Each agent has an Actor-Critic structure, and the Actor network parameters can be expressed as The critic network parameters can be expressed as The agent determines the optimal migration strategy by maximizing the reward. However, in the interaction between the agent and the environment, the recent reward is more important than the future reward. Therefore, there is a discount factor γ∈[0,1), and the global cumulative discounted reward Let the global state value function be V π (s t ) indicates state s t The long-term total reward under , where π is the strategy, and the state value function is expressed as:

[0248]

[0249] in is the expected operator, then the optimal VNF migration strategy π * It can be expressed as:

[0250]

[0251] The Actor network updates the network parameters through the policy loss function. The policy loss function of the Actor network corresponding to agent k can be expressed as:

[0252]

[0253] in, is the ratio of the old and new strategies of the Actor network, is the advantage function, and clip() is a truncation function to ensure the update amplitude of the new and old parameters. A hyperparameter for setting the clipping range of the clip() function.

[0254] The parameter update of the critic network is based on the state value loss function, which is defined as the mean square error between the cumulative discounted reward and the state value function, which can be expressed as:

[0255]

[0256] in, is the cumulative discounted reward of agent k, is the state value function of agent k.

[0257] Finally, this embodiment uses the gradient ascent method to update the parameters of the Actor network and the gradient descent method to update the parameters of the Critic network, which can be expressed as:

[0258]

[0259]

[0260] Among them, α a and α c are the learning rates of the Actor network and the Critic network respectively.

[0261] After the parameters are updated, the Actor network and the Critic network will improve the strategy and state value function according to the new parameters, thereby enhancing the performance of the agent in the environment. With the continuous iteration of training, the parameters of the strategy and value function will gradually stabilize, the overall performance of the system will be improved, and ultimately the VNF migration strategy will be optimized.

[0262] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.

Claims

1. A digital twin-assisted virtual network function resource prediction and migration method, characterized in that: It includes the following steps: S1. Build a digital twin-assisted VNF resource prediction and migration system model, which includes the physical infrastructure layer, virtual network layer, and digital twin network layer; S2. Establish an end-to-end delay model, system energy consumption model, cost model, load balancing model, information age violation probability model, and the optimization goal of maximizing system QoS based on the VNF resource prediction and migration system model; S3. Establish an AMT-DMG spatiotemporal fusion prediction model to predict the resource requirements of VFN, wherein the spatiotemporal features are extracted and fused by the attention-based multi-scale temporal convolutional network TCN and the distance-based multi-head graph attention network GAT to obtain the resource demand prediction results of VNF; during the training process, the credibility-weighted group learning framework is embedded in the AMT-DMG spatiotemporal fusion prediction model, and the convergence of the training model is accelerated by sharing the training parameters of the AMT-DMG spatiotemporal fusion prediction model at each VDT node; S4. Based on the prediction results of VNF resource requirements, a multi-agent proximal strategy optimization algorithm is used to solve the migration strategy with maximizing system QoS as the optimization goal.

2. According to claim 1, a digital twin-assisted virtual network function resource prediction and migration method is characterized in that: In step S1, the physical infrastructure layer provides physical network resources for the virtual network layer. The physical infrastructure layer is modeled as a connected undirected graph in is the underlying physical node set, is the underlying physical link set, and the CPU resources and storage resources of physical node n are and represents the physical link between physical nodes n and m, and physical link l nm The bandwidth resource is The physical nodes include twin service nodes for building VNF virtual service nodes and twin service nodes for building VNF digital twins; The virtual network layer is an undirected graph structure network consisting of virtual nodes mapped by multiple physical devices and virtual links mapped by physical links. It builds a service function chain SFC according to the business requests of different users. Suppose a set of SFCs is represented as Each SFC is a set of VNFs arranged in order, where represents the i-th SFC, represents the VNF set of the i-th SFC, represents the jth VNF on the i-th SFC, for The virtual link set is Abstract as a directed graph The CPU resource requirements and storage resource requirements are and Virtual Link express and Virtual link between The bandwidth resource requirement is Defining binary variables Used to reflect the mapping relationship between virtual network and physical network. When At time t, it is mapped to the physical node n. When At time t, it is not mapped to physical node n; when When , it indicates a virtual link Mapped to physical link l at time t nm On, when When , it indicates a virtual link At time t, there is no mapping to the physical link l nm superior; The digital twin network layer builds a digital twin network of the virtual network by collecting valid status information in the physical network and the virtual network. Each VNF node will be associated with a digital twin service node and a corresponding digital twin will be built. The digital twin of the VNF is a VDT node. The future resource demand results of the VDT node are predicted based on the historical resource data in the data collection library. This is used to determine whether the VNF node will have an abnormal situation at the next moment, formulate a VNF migration strategy in advance, and then send the strategy to the physical network through the SDN controller for implementation.

3. According to claim 2, a digital twin-assisted virtual network function resource prediction and migration method is characterized in that: In step S2, for the end-to-end delay model, an SFC service request The end-to-end delay τ at time t total (t) is expressed as: in, yes size; is the number of network routing hops between physical node m and physical node n; t is the delay required to transmit one unit of data along the migration path, that is, the transmission rate of VNF migration; Θ is a positive coefficient; Indicates physical link l nm The physical distance; c is the speed of light; express The processing rate; is the processing rate coefficient of the data packet; Service Request for SFC The packet size; For the system energy consumption model, the total energy consumption of the network at time t is P total (t) is expressed as: Among them, the energy consumption of the underlying physical node n includes the energy consumption in the active state Energy consumption in sleep mode and the energy consumption of switching states B n (t) is the switch binary variable of the physical node, B n (t) = 1 means the node is turned on, Z n (t) is a binary variable indicating whether the state of the physical node n has changed at time t, expressed as: Among them, Z n (t) = 1 means that the state of the management node n changes at time t, Z n (t) = 0 means that the state of the management node n at time t has not changed; For the cost model, when the VNF instance carried by a physical node needs to be migrated to another physical node that can meet the VNF resource requirements, at time t, the overall network migration cost is expressed as: Among them, κ is the migration cost factor, Is a node The size of is the number of network routing hops between physical node m and physical node n; For the load balancing model, the variance of the CPU resource load, storage resource load, and link bandwidth resource load in the network can be expressed as: in, and Indicates the CPU resource and storage resource load of physical node n at time t, For physical link bandwidth resource load, and It represents the average value of the system's CPU resource load, storage resource load, and link bandwidth resource load at time t, and the load balancing capability of the network system L total (t) is expressed as: Among them, μ1=μ2=μ3=1 / 3, that is, CPU resources, storage resources and bandwidth resources have the same impact on network load balancing; For the information age violation probability model, assume that the generation time of the Mth data packet of the VDT synchronization information is t M , the time when the VDT node receives the update information packet is t' M , X M =t M -t M-1 is the time interval between two adjacent data packets, T M =t' M -t M is the time the data packet spends in the system, Y M =t' M -t' M-1 is the time interval between two adjacent data packets arriving at the VDT node, is the AoI constraint threshold, then at time t, the instantaneous information age Δ(t), t>0 is expressed as: Δ(t)=t-max{t M :t' M ≤t} Among them, max{t M :t' M ≤t} is the timestamp of the most recent update information received by the VDT node before time t; G(M) indicates that the AoI of the Mth synchronization data packet is greater than the threshold constraint The duration of G(M) is expressed as: The AoI violation probability is defined as the AoI of the VDT node successfully receiving the update information and the AoI exceeding the given AoI threshold. The probability is expressed as: in, The number of times the VDT node receives update information packets.

4. According to claim 3, a digital twin-assisted virtual network function resource prediction and migration method is characterized in that: In step S2, the QoS of the system is expressed as: Q=ξ1Q Cost +ξ2Q Power +ξ3Q Load Among them, Q Cost , Q Power and Q Load They represent the QoS of system migration cost, the QoS of system energy consumption, and the QoS of system load balancing, respectively. ξ1, ξ2, and ξ3 are the weight factors of migration cost, energy consumption, and load balancing indicators on system QoS. ξ1+ξ2+ξ3=1 and satisfy ξ1,ξ2,ξ3∈(0,1). The optimization objective is subject to the following constraints: C4:t total (t)≤τ req Among them, C1 ensures that each VNF is mapped to only one physical node; C2 ensures that each virtual link is mapped to only one physical link; C3 represents the AoI violation probability limit of the twin synchronization data, Pr max is the maximum AoI violation probability; C4 represents the end-to-end delay limit of SFC, Request for SFC The end-to-end delay requirement is 100%. C5 indicates that the physical node carrying the VNF is in the turned-on state. C6-C8 indicate that the CPU resources and storage resources of the VNF cannot exceed the resource limits of the physical node, and the virtual link cannot exceed the bandwidth resource limit of the physical link. and are the upper and lower limit factors of resource utilization of physical network resources; C9-C10 indicate that when a virtual link is mapped to a physical link, the corresponding VNF node must also be mapped to the corresponding physical node; C11-14 are binary variable constraints during VNF migration.

5. According to claim 1, a digital twin-assisted virtual network function resource prediction and migration method is characterized in that: In step S3, the input time series data is x={x(t)|t=1,2,...,T′}, T′ is the total length of the time step of the data, and the feature data is x(t)={x j (t)|j=1,2,...,N′}, N′ is the total number of VDT nodes, x j (t) is the characteristic data of the jth VNF, and finally the characteristic data of the VNF at time t+1 is output x(t+1)={x j (t+1)|j=1,2,...,N′}; Among them, the process of extracting temporal features by the attention-based multi-scale temporal convolutional network TCN model is: The TCN model includes TCN residual blocks with filters of different sizes, each TCN residual block includes two dilated causal convolutions and weight normalization, and outputs the results through residual connections and ReLU activation functions, where the calculation formula for the dilated causal convolution is: In the formula, x(t) is the input time series data feature, For the The output of the dilated causal convolution in the TCN, For the The convolution kernel size of a TCN, f(g) is the g-th element in the convolution kernel, x(td g) represents the data at different time steps, and d is the dilation coefficient; The output of the residual connection is the temporal features extracted by the TCN network at different time scales. The calculation formula is as follows: The calculation process of the time scale attention layer is as follows: In the formula, w1 and w2 are weight coefficients, b is the bias coefficient, To normalize the attention weight value, the output of the attention layer is used express; The results of the TCN networks at different time scales after importance weighting are spliced, and then input into the average pooling layer and the fully connected layer to obtain the time feature matrix H t : Among them, concat(·) represents the concatenation function, Avgpooling(·) is the average pooling layer operation, and FC(·) is the fully connected layer operation.

6. A digital twin-assisted virtual network function resource prediction and migration method according to claim 5, characterized in that: In step S3, the distance-based multi-head graph attention network GAT model extracts spatial features as follows: Define the node feature input vector as: in, is the feature representation of node u, is the feature vector dimension of the node, U is the number of VNF nodes in the network; The state feature update of each node is expressed as: Among them, δ(*) is a nonlinear activation function, and the feature representation of the central node u is updated through aggregation operation as follows: Λ u Represents the label set of neighbor nodes of node u, Φ' uu′ is the attention coefficient, expressed as: Among them, ε is the attention coefficient threshold, d u ' u′ is the reciprocal normalized distance, Φ uu′ is the normalized attention coefficient of the influence of neighbor node u′ on the central node u; Introduce a multi-head attention mechanism to focus on multiple different features of the input. Suppose the number of attention heads of the attention mechanism is K head , then the updated feature representation of the central node is: Aggregate the update results of each node to obtain the spatial feature matrix of the VNF node The time feature matrix H t With the spatial feature matrix H s The input is integrated into the fully connected layer for fusion and dimension reduction to obtain the spatiotemporal feature matrix representing the VNF resource demand characteristics, and then the prediction result of the VNF resource demand at the next moment is obtained: H=FC2(FC1(H t ,H s )) H={x j (t+1)|j=1,2,...,N′} Where H = FC2(FC1(H t ,H s )) is the fused spatiotemporal feature matrix, FC1 and FC2 are different fully connected layers; H = {x j (t+1)|j=1,2,...,N′} is the prediction result of VNF resource demand at the next moment.

7. According to claim 5, a digital twin-assisted virtual network function resource prediction and migration method is characterized in that: In step S3, the process of embedding the credibility-weighted group learning framework in the AMT-DMG spatiotemporal fusion prediction model is as follows: 1) Each VDT node has a local AMT-DMG spatiotemporal fusion prediction model. First, each VDT node runs its own AMT-D spatiotemporal fusion prediction model locally and uploads its own local model training parameters to the SL network: The SL network initializes the global model parameters as θ, and the jth VDT node combines its local sample data set To train local model parameters Use the gradient descent algorithm to update the local training model parameters: Among them, λ′>0 is the learning rate, is the loss function over parameters The gradient of 2) Secondly, the SL network selects a temporary central integration node, which uses a credibility weighted algorithm to merge the model training parameters of all participants: The credibility value of each VDT is expressed as is the total number of VDT nodes, and the selection process of the temporary central integration point is expressed as: After uploading the model parameters of all participants to the temporary central integration node, the model parameters are merged and the global model parameter θ is obtained using the weighted average function. o , expressed as: 3) Finally, the temporary central integration node distributes the merged model parameters to other participants for training and dynamically updates the local model, repeating the training until the entire training process converges.

8. The method for predicting and migrating virtual network function resources assisted by digital twins according to claim 1, characterized in that: In step S4, the following steps are included: S41: Calculate resource utilization of CPU resources, storage resources, and bandwidth resources of physical nodes based on the resource prediction results of the VNF, and select overloaded or underloaded VNF nodes that need to be migrated; S42: For VNF migration, the system QoS is optimized to the maximum extent possible under the conditions of meeting the end-to-end delay and the age limit of the digital twin synchronization data information. The migration cost, network energy consumption, and network load balancing are used to measure the system QoS. S43: A multi-agent proximal strategy optimization deep reinforcement learning algorithm is used to solve the VNF migration problem. Each SFC is regarded as an agent, and the optimization goal is to maximize the global reward of the system to obtain the optimal VNF migration strategy.

9. A digital twin-assisted virtual network function resource prediction and migration method according to claim 8, characterized in that: In step S43, the process of solving the migration strategy by the multi-agent proximal strategy optimization algorithm is as follows: Suppose there are K SFC requests, each SFC is defined as an agent, and the MDP model of each agent is defined as Represent the state space Action Space State transition probability Reward Function Then the global reward of the multi-agent system is It is expressed as: in, represents the instantaneous reward of agent k; Each agent has an Actor-Critic structure, and the Actor network parameters are expressed as The critic network parameters are expressed as They represent the Actor network parameters and Critic network parameters of agent k respectively; Introducing a discount factor γ∈[0,1), a global cumulative discount reward Among them, i represents the number of steps from the current time step to the future. represents the global reward at time t+i; Assume that the global state value function is Vπ(s t ) indicates state s t The long-term total reward under , where π is the strategy, and the state value function is expressed as: in is the expected operator, then the optimal VNF migration strategy π * It is expressed as: The Actor network updates the network parameters through the policy loss function. The policy loss function of the Actor network corresponding to agent k is expressed as: in, is the ratio of the old and new strategies of the Actor network, is an advantage function, clip() is a truncation function, To set the hyperparameters for clipping range of the clip() function; The parameter update of the critic network is based on the state value loss function, which is defined as the mean square error between the cumulative discounted reward and the state value function, which is expressed as: in, is the cumulative discounted reward of agent k, is the state value function of agent k; Finally, the gradient ascent method is used to update the parameters of the Actor network, and the gradient descent method is used to update the parameters of the Critic network, which are expressed as: Among them, α a and α c They are the learning rates of the Actor network and the Critic network respectively; After the parameters are updated, the Actor network and the Critic network will improve the strategy and state value function according to the new parameters, and continuously iterate the training to optimize the VNF migration strategy.

10. A digital twin-assisted virtual network function resource prediction and migration method according to claim 9, characterized in that: For each agent in the MDP model, the state space There are: The state space of the kth agent is defined as in represents the network state space of physical nodes, represents the state space of the physical link, and Represents the computing resource requirement set and storage resource requirement set of the VNF in an SFC. is a set of virtual link bandwidth resource requirements; For each agent's action space in the MDP model There are: define the action space of the kth agent in A collection of actions mapped to virtual nodes, is a set of virtual link mapping actions; For each agent's MDP model, the state transition probability function There are: The state transition probability function represents the agent in state s t Take action a t , and transfer to a new state at the next moment The conditional probability distribution of t ×a t ×s t+1 →[0,1], a set is used to represent the conditional probability distribution, which contains the probability of each action taken in the current state to transfer to the next state. The state transition probability is expressed as For each agent's MDP model, the reward function At time t, the kth agent is in state Take Action Will generate a strategy According to the optimization objectives and constraints of the VNF migration problem, the instantaneous reward of agent k is obtained for: Among them, Q Cost , Q Power and Q Load They represent the QoS of system migration cost, the QoS of system energy consumption and the QoS of system load balancing respectively. ξ1, ξ2 and ξ3 are the weight factors of migration cost, energy consumption and load balancing indicators on system QoS. Ω represents the penalty factor.

Citation Information

Patent Citations

  • Automated resource management for distributed computing

    CN113795826A

  • Digital twinning auxiliary radio resource allocation method based on time sequence residual stacking network

    CN118283659A

  • Optimization of deterministic and non-deterministic traffic in radio-access network (RAN)

    US20240057139A1

Cited By

  • Distributed real-time database intelligent management method and system based on big data

    CN119356882A

  • Service chain determination method and device, equipment and storage medium

    CN119496707A

  • Digital twinning-based multi-task learning petrochemical device anomaly positioning method

    CN119691643A

  • VNF multi-dimensional resource demand prediction method based on synchronous and asynchronous space-time characteristic fusion

    CN119740095A

  • Multi-scale digital twinning system of combustion explosive production line

    CN119987307A