A network slice resource prediction method and system based on ring federated learning

Abstract

The present invention discloses a network slice resource prediction method and system based on ring federated learning, comprising: obtaining performance indicator data required by each VNF in a VNF sequence for performing network slice services within a continuous time; obtaining a prediction result of the resources required for the network slice based on the performance indicator data required by each VNF for performing network slice services within a continuous time and a trained network slice resource prediction model; wherein the network slice resource prediction model is constructed using a ring federated learning framework, treating the physical node where each VNF is located as a client participating in federated learning, constructing a node resource prediction model for each client, inputting the performance indicator data required by each VNF for performing network slice services within a continuous time into the node resource prediction model of the physical node where the VNF is located, and outputting a prediction result of the resources required by the network slice at the next moment. This improves the accuracy of network slice resource prediction.

Description

Technical Field

[0001] The present invention relates to the field of network slice resource prediction technology, and in particular to a network slice resource prediction method and system based on ring federated learning. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] As a key enabling technology for 5G and future mobile communication networks, network slicing provides users with customized services by creating multiple isolated and customized virtual networks that meet different service requirements on a public physical infrastructure network, solving the problem that traditional networks are difficult to meet diverse performance requirements. However, in the network slicing scenario, the traditional network method of fixed resource allocation may fail to meet different performance requirements through fixed and unreasonable resource allocation.

[0004] To this end, a method for predicting the resources required for network slicing through machine learning algorithms has been developed. However, when existing machine learning algorithms predict the resources required for network slicing, they aggregate the performance indicator data observed by all VNFs to train the network slicing resource optimization model, which not only leads to huge communication overhead but also violates the requirements of data security and privacy protection. Summary of the Invention

[0005] In order to solve the above problems, the present invention proposes a network slice resource prediction method and system based on ring federated learning. When predicting the resources required for network slicing, it can not only ensure the data security between each VNF, but also perform data fusion value mining, making the prediction of the resources required for network slicing more accurate.

[0006] To achieve the above object, the present invention adopts the following technical solutions:

[0007] First, a network slice resource prediction method based on ring federated learning is proposed, including:

[0008] Obtain the performance indicator data required for each VNF in the VNFs sequence to provide network slicing services over a continuous period of time;

[0009] Based on the performance indicator data required by each VNF in the VNFs sequence to provide network slice services in a continuous time and the trained network slice resource prediction model, the prediction result of the resources required by the network slice at the next moment is obtained;

[0010] Among them, the network slicing resource prediction model is constructed using a circular federated learning framework. The physical node where each VNF is located is regarded as a client participating in federated learning. A node resource prediction model is built for each client. The performance indicator data required for each VNF to provide network slicing services in a continuous time is input into the node resource prediction model of the physical node where the VNF is located, and the resource prediction results required for the network slice at the next moment are output.

[0011] Secondly, a network slice resource prediction system based on ring federated learning is proposed, including:

[0012] The data acquisition module is used to obtain the performance indicator data required by each VNF in the VNFs sequence to provide network slicing services in a continuous time;

[0013] The network slice required resource prediction module is used to obtain the prediction result of the resources required by the network slice at the next moment based on the performance indicator data required by each VNF in the VNFs sequence to provide network slice services in a continuous time and the trained network slice resource prediction model;

[0014] Among them, the network slicing resource prediction model is constructed using a circular federated learning framework. The physical node where each VNF is located is regarded as a client participating in federated learning. A node resource prediction model is built for each client. The performance indicator data required for each VNF to provide network slicing services in a continuous time is input into the node resource prediction model of the physical node where the VNF is located, and the resource prediction results required for the network slice at the next moment are output.

[0015] On the third aspect, an electronic device is proposed, including a memory and a processor, and computer instructions stored in the memory and running on the processor. When the computer instructions are executed by the processor, the steps of a network slice resource prediction method based on ring federated learning are completed.

[0016] In a fourth aspect, a computer-readable storage medium is proposed for storing computer instructions. When the computer instructions are executed by a processor, the steps of a network slice resource prediction method based on ring federated learning are completed.

[0017] Compared with the prior art, the present invention has the following beneficial effects:

[0018] 1. The present invention adopts a ring federated learning framework to construct a network slice resource prediction model. The physical node where each VNF is located is regarded as a client participating in federated learning. A node resource prediction model is constructed for each client. When the network slice resource prediction model is trained, the node resource prediction model parameters trained on each node are sequentially transmitted to the next node as the global model parameters for the training of the next node. This can not only fully utilize the fusion value between each VNF, but also ensure the independent security of each VNF. When the trained network slice resource prediction model is used to predict the resources required for the network slice, it can not only ensure the data security between each VNF, but also mine the fusion value of each VNF, and make the prediction of the resources required for the network slice more accurate, thus solving the problem of poor performance of the network slice resource optimization model caused by the data island problem.

[0019] 2. When training the network slice resource prediction model, this invention uses a triplet loss to constrain updates to the client node resource prediction model (local model). This aligns the time series features extracted by the local model with those extracted by the global model, training the local model closer to the global model. This allows for the acquisition of more data fusion information, further ensuring the accuracy of resource predictions required for network slicing.

[0020] Advantages of additional aspects of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The drawings in the specification, which constitute a part of this application, are used to provide further understanding of this application. The illustrative embodiments of this application and their descriptions are used to explain this application and do not constitute improper limitations on this application.

[0022] Figure 1 This is the training process of the model in the method disclosed in Example 1;

[0023] Figure 2 The ring federated learning topology disclosed in Example 1;

[0024] Figure 3 This is a prediction flow chart of the node resource prediction model disclosed in Example 1;

[0025] Figure 4 This is a flow chart of the long short-term memory network unit disclosed in Example 1. DETAILED DESCRIPTION

[0026] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0027] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present application belongs.

[0028] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0029] Example 1

[0030] Network slicing, a key enabling technology for 5G and future mobile communication networks, provides users with customized services by creating multiple, isolated, customized virtual networks that meet different service requirements on a common physical infrastructure. This addresses the difficulty traditional networks face in meeting diverse performance requirements. However, in network slicing scenarios, fixed and irrational resource allocation may not be able to meet diverse performance requirements. Therefore, this embodiment proposes a network slice resource prediction method based on ring federated learning to maximize the performance requirements of the slice at different times. Furthermore, the method disclosed in this embodiment can optimize the network slice resource prediction model while protecting data security.

[0031] As networks with capabilities such as massive access, deterministic latency, and extremely high reliability, 5G networks should be able to provide flexible and dynamic networks for a wide range of users to meet the diverse business needs of users and vertical industries. Therefore, traditional approaches to fixed resource allocation are no longer effective. Network slicing, a new network architecture, leverages Network Function Virtualization (NFV) and Software-Defined Networking (SDN) technologies to enable network operators to provide a variety of customized services to users in a sustainable manner. By leveraging NFV to decouple hardware and software, and SDN to decouple control and data, network slicing can provide multiple independent logical networks on a shared network infrastructure. Each logical network serves a specific business type or industry user, while providing highly flexible, on-demand network services based on business needs. Furthermore, network slices based on NFV and SDN can flexibly define their own logical topology, reliability, and security levels to meet the differentiated needs of different businesses, industries, or users. The flexibility of network slicing can also meet the growing business needs of the future. Through the powerful learning ability of artificial intelligence technology, network slice resources can be further managed intelligently, enabling network slices to adapt to dynamic changes in the network environment, further improving the business performance of slices, and saving operating costs while optimizing communication service quality.

[0032] In a network slicing architecture based on SDN / NFV technologies, user services within a network slice are served by a service function chain (SFC). An SFC is composed of multiple virtual network functions (VNFs) in sequence. When a service is provided, data flows sequentially through all the VNFs in the SFC. By leveraging virtualization technology, a series of VNFs are decoupled from dedicated network functions. The VNFs required for communication are mapped onto physical nodes in an efficient and flexible manner to provide customized services. Physical nodes allocate resources to VNFs to support their operation and complete service provision. Therefore, resource allocation within a slice is essentially the allocation of resources across a series of VNFs. In real-world network environments, the resources required for a slice can also change dynamically. Static resource allocation methods are not suitable for dynamically changing network conditions. Therefore, developing efficient and dynamic slice resource management strategies to address these dynamic network conditions has attracted widespread attention.

[0033] Because the VNFs in a service function chain (SFC) are arranged in sequence, and data flows sequentially through all VNFs in the chain during service delivery, even in dynamically changing scenarios, the resources required by each VNF are somewhat similar. By monitoring the performance metrics required by VNFs in a network slice over time, the resource requirements for the network slice can be predicted, allowing resources to be allocated on demand. However, aggregating the performance metrics observed by VNFs to train the network slice resource optimization model not only incurs significant communication overhead but also violates data security and privacy requirements. Therefore, federated learning technology is employed to perform data fusion value mining while ensuring data security and data immobility, addressing the data silo issue caused by security and privacy. Federated learning, as a distributed machine learning framework, can be categorized into star, ring, and peer-to-peer topologies. This embodiment, based on the unique chain structure of VNFs in a network slice, selects a ring-based federated learning architecture. Each physical node where a VNF resides is treated as a client participating in federated learning training. A continuous ring connects each node to train the network slice resource optimization model.

[0034] A network slice resource prediction method based on ring federated learning disclosed in this embodiment is described in detail.

[0035] A network slice resource prediction method based on ring federated learning, comprising:

[0036] S1: Obtain the performance indicator data required for each VNF in the VNFs sequence to provide network slicing services continuously over time.

[0037] Since each service function chain (SFC) in a network slice is composed of multiple virtual network functions (VNFs) in an orderly manner, when a user service arrives, a series of VNFs will provide services to the user, and the service process will pass through all VNFs in the SFC in sequence. In order to predict the resources required for the network slice, we first obtain the VNF sequence when the user service arrives, and obtain the performance indicator data required for each VNF in the VNF sequence to provide network slice services in a continuous time.

[0038] S2: Based on the performance indicator data required by each VNF in the VNFs sequence to provide network slice services in a continuous time and the trained network slice resource prediction model, the prediction result of the resources required by the network slice at the next moment is obtained;

[0039] Among them, the network slicing resource prediction model is constructed using a circular federated learning framework. The physical node where each VNF is located is regarded as a client participating in federated learning. A node resource prediction model is built for each client. The performance indicator data required for each VNF to provide network slicing services in a continuous time is input into the node resource prediction model of the physical node where the VNF is located, and the resource prediction results required for the network slice at the next moment are output.

[0040] Based on the special structure of the service function chain SFC, this embodiment adopts a ring federated learning framework to build a network slice resource prediction model. Ring federated learning is a decentralized federated learning structure that uses a continuous ring to connect each node participating in the training. In this embodiment, the physical node where each VNF on the service function chain is located is regarded as a client participating in federated learning. A local model is built for each client. The local model adopts the node resource prediction model, and the overall collaborative learning model between the physical nodes is used as the global model.

[0041] Get the local VNF performance indicator dataset of each client. The local VNF performance indicator dataset includes the performance data X={x1,x2,...,x t}, input the local VNF performance indicator data set into the node resource prediction model of the corresponding client, and train the network slice resource prediction model. During training, the node resource prediction model parameters trained on each node are sequentially transferred to the next node as the global model parameters for the next node training. The local VNF performance data on the next node is used to train the local model of the node until the specified iteration rounds or the set target accuracy are met, and the training is stopped to obtain the trained network slice resource prediction model. The structure of the ring federated learning is as follows: Figure 2 shown.

[0042] Among them, the node resource prediction model, such as Figure 3 As shown, it includes CNN and LSTM networks. CNN extracts performance indicator features from performance indicator data, splices the performance indicator features according to the time dimension to form an indicator sequence, and inputs the indicator sequence into the LSTM network to obtain the prediction result of the resources required by the network slice at the next moment.

[0043] When VNFs in a network slice provide services to user businesses, multiple performance indicators are often involved, including resource data such as CPU, bandwidth, and memory required for VNFs operation. There are also correlations between different performance data. Therefore, this embodiment uses a convolutional neural network (CNN) to capture the interaction between various performance indicators. That is, for each physical node where the VNF is located, a convolutional neural network (CNN) is used to extract the spatial relationship between sample features at the tth moment. The performance indicator data of the VNF in the network slice at the tth moment is represented as a vector n represents the sample x t The feature dimension of is n-dimensional. The local dependency relationship of the feature space is extracted in the spatial dimension by the convolutional neural network CNN to obtain the performance index characteristics. The sample after extracting the spatial relationship can be expressed as The sample f t The feature dimension is m.

[0044] As a recursive neural network, the long short-term memory network (LSTM network) is suitable for processing time series data and can solve the gradient vanishing or gradient exploding problems of the original recurrent neural network. The LSTM network can predict the resources required for the network slice at time (t+1) based on the data of the previous t moments in history, thereby allocating network resources reasonably and on demand to achieve the purpose of optimizing network slice resources. In addition, the LSTM network has memory to overcome the long-term time dependency problem associated with the input sequence, and can effectively express the information in the long time series without forgetting the useful information from a long time ago. The structure of a single LSTM unit is as follows: Figure 4 As shown, there are three gates, namely input gate, output gate, and forget gate, as well as two types of memory: long-term memory and short-term memory.

[0045] The formula of the forget gate is as follows, f t is the output of the forget gate, W f and b f are the network weight and bias term of the forget gate, h t-1 is the hidden layer output at time t-1, x t is the feature vector input at time t, and σ is the activation function:

[0046] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0047] The formula of the input gate is as follows, where i t is the output of the input gate, W i is the network weight of the input gate, b i is the bias term of the input gate.

[0048] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0049] Then the main line memory in LSTM is replenished, that is The formula is as follows, where the activation function uses the tanh function and the input gate controls the replenishment size:

[0050]

[0051] By forgetting the main line memory through the forget gate and replenishing the main line memory through the input gate, the main line memory can be updated and the long-term memory and short-term memory can be integrated. The formula is as follows:

[0052]

[0053] Among them, C t-1 is the “memory” of the recurrent neural network for all input information at the previous t-1 moment, C t It is the “summary” of the input data before t moments after passing through the t-th LSTM unit.

[0054] The formula of the output gate is as follows, where o t is the output of the output gate, W o is the network weight of the output gate, b o is the bias term of the output gate.

[0055] o t =σ(W o ·[h t-1 ,x t ]+b o )

[0056] The output o of the output gate t With mainline memory C t Combined, we can get the hidden layer output h of this LSTM unit t , the formula is as follows, where h t It is also the input of the next LSTM unit.

[0057] h t =o t *tanh(C t )

[0058] By fusing long-term and short-term memory, the memory in the neural network is dynamically processed, allowing the neural network to learn which information needs to be remembered long-term and which memories can be forgotten. This embodiment uses the long-term and short-term memory network to predict the performance resources required for network slicing, thereby allocating resources reasonably. Assume that the performance data of the VNF on the physical node in t time periods is represented by X = {x1, x2, ..., x t}, X is input into the LSTM unit as the feature vector of the previous t time points, and x can be predicted. t+1 The performance indicators required by the VNF in the network slice at all times are used to achieve resource optimization by making reasonable predictions on the network slice resources.

[0059] Therefore, the performance index features of the CNN feature extractor are spliced ​​according to the time dimension to obtain the index sequence F, F = {f2, f1, ..., f t Without changing the time characteristics of the input sequence, it is sequentially input into the LSTM unit in the long short-term memory network (LSTM network) to predict the resources required for the network slice at time t+1. After going through t LSTM units, the hidden layer output h of the tth LSTM unit is obtained. t , adding a fully connected layer after it can serve as the actual output of the LSTM network and predict the performance resources required by the VNF at time t+1.

[0060] For each VNF node in the service function chain (SFC), a node resource prediction model is used to predict the resources required for the network slice. The input time series data is first extracted using a convolutional neural network, and then a LSTM network is used for sequence prediction to predict the performance resources required for the network slice.

[0061] During the operation of the network slice, when user services arrive, the VNFs on the service function chain will serve the user in sequence, and the data flow will pass through all VNFs in the SFC completely. Therefore, the performance data generated by each VNF is similar to a large extent. However, sharing the data samples generated by multiple VNFs and jointly training a network slice resource optimization model will violate data security and privacy requirements. In order to alleviate the data island problem while protecting data security, this embodiment uses a federated learning framework. Federated learning, as a distributed machine learning framework, can use data in a legal manner while protecting data privacy.

[0062] The local VNF performance data of each node is input into the node resource prediction model of the corresponding node, and the network slice resource prediction model is trained. During training, it is assumed that the node resource prediction model parameters of the i-th client in the r-th round are In the r+1th round, the previous client j on the i-th client topology will train the node resource prediction model parameters Passed to the i-th client, client i uses it as the global model parameter of the r+1th round and The node resource prediction model of client i is trained using the local VNF performance data of client i to obtain the node resource prediction model parameters updated in round r+1.

[0063] When training the network slice resource prediction model, the triplet loss is used to constrain the update direction of the node resource prediction model at the node. The loss function is calculated by the difference between the feature representations learned by the local model and the global model, and the global model is used to constrain the update of the local model. Each time the local training is performed, the VNF performance data X={x1,x2,...,x t} as a set of local VNF performance data and fed into three model parameters: In the example, we can get the local model parameters of the i-th client in the r-th round The calculated hidden layer output at time t is The hidden layer output at time t calculated by the global model on the i-th client in the r+1th round is Updated local model The calculated hidden layer output at time t is The triplet loss is applied between these three feature representations, and the formula is as follows:

[0064]

[0065] Among them, d represents the Euclidean distance, margin represents the threshold, and the triple loss L tl , constraining the local model to train in a direction closer to the global model.

[0066] When training the network slice resource prediction model, the cross entropy loss is calculated for the node resource prediction model, and the cross entropy loss is used to constrain the training of the node resource prediction model.

[0067] In addition, when the local model of the i-th client is trained in the r+1 round, the time series samples in the t time period are input into the node resource prediction model each time, and the hidden layer features output by the t-th long short-term time memory network unit are obtained. The hidden layer features output by the t-th long short-term time memory network unit are passed through a fully connected layer to obtain the prediction of the resources required for the network slice at time t+1. The predicted value is compared with the actual value of the performance indicator required at time t+1. The cross entropy loss L is calculated. c The calculation of , thereby further constraining the model training. The formula of the overall loss function L of the model is as follows:

[0068] L=Lc +L t1 .

[0069] When training the network slice resource prediction model, the stochastic gradient descent method is used to update the parameters of the node resource prediction model.

[0070] During local model training, the calculated loss value is updated using stochastic gradient descent. When the pre-set number of federated learning iterations or the target effect is achieved, a federated learning-based network slice resource prediction model is learned. When the network slice resource optimization model learned in this embodiment is used to predict network slice resources, the performance indicator data of the VNF in the previous t time periods is input. The performance indicator of the VNF on the physical node at time t+1 can be predicted, thereby reasonably predicting resources and allocating them on demand, achieving network slice resource optimization.

[0071] This embodiment combines the structure of the service function chain in the network slicing architecture and adopts a federated learning framework with a ring topology. The decentralized ring federated learning structure is suitable for network slicing usage scenarios and can simply and effectively coordinate various VNFs. The ring federated learning structure designed in this embodiment can jointly train the model while protecting the data security and privacy of each participating client, solving the problem of suboptimal performance of the network slice resource optimization model caused by data silos. This embodiment uses a triplet loss to constrain the update of the client's local model. By aligning the temporal features extracted by the local model with the temporal features extracted by the global model, the local model is trained in a direction closer to the global model. During local training on each client, this embodiment adopts a CNN+LSTM network architecture. The input training samples are first subjected to spatial feature extraction using a CNN feature extractor. While maintaining the time series, the LSTM is then used to extract temporal features to predict network slice resources. The network slice resource optimization method learned in this patent can dynamically adjust and allocate network slice resources on demand in scenarios where the network is dynamically changing.

[0072] Example 2

[0073] In this embodiment, a network slice resource prediction system based on ring federated learning is disclosed, including:

[0074] The data acquisition module is used to obtain the performance indicator data required by each VNF in the VNFs sequence to provide network slicing services in a continuous time;

[0075] The network slice required resource prediction module is used to obtain the prediction result of the resources required by the network slice at the next moment based on the performance indicator data required by each VNF in the VNFs sequence to provide network slice services in a continuous time and the trained network slice resource prediction model;

[0076] Among them, the network slicing resource prediction model is constructed using a circular federated learning framework. The physical node where each VNF is located is regarded as a client participating in federated learning. A node resource prediction model is built for each client. The performance indicator data required for each VNF to provide network slicing services in a continuous time is input into the node resource prediction model of the physical node where the VNF is located, and the resource prediction results required for the network slice at the next moment are output.

[0077] Example 3

[0078] In this embodiment, an electronic device is disclosed, including a memory and a processor, and computer instructions stored in the memory and running on the processor. When the computer instructions are executed by the processor, the steps described in the network slice resource prediction method based on ring federated learning disclosed in Example 1 are completed.

[0079] Example 4

[0080] In this embodiment, a computer-readable storage medium is disclosed for storing computer instructions. When the computer instructions are executed by a processor, the steps described in a network slice resource prediction method based on ring federated learning disclosed in Example 1 are completed.

[0081] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered by the scope of protection of the claims of the present invention.

Claims

1. A network slice resource prediction method based on ring federated learning, characterized in that: include: Obtain the performance indicator data required for each VNF in the VNFs sequence to provide network slicing services over a continuous period of time; According to the performance indicator data required by each VNF in the VNFs sequence to provide network slice services in a continuous time and the trained network slice resource prediction model, the prediction result of the resources required by the network slice at the next moment is obtained. When training the network slice resource prediction model, the update direction of the node resource prediction model is constrained by the triplet loss. The triplet loss L tl for: Among them, d represents the Euclidean distance, margin represents the threshold, and the node resource prediction model parameters of the i-th client in the r-th round are The calculated hidden layer output at time t is The hidden layer output at time t calculated by the global model on the i-th client in the r+1th round is Updated node resource prediction model The calculated hidden layer output at time t is The network slice resource prediction model is built using a circular federated learning framework. Each physical node where the VNF is located is treated as a client participating in federated learning. A node resource prediction model is built for each client. The performance indicator data required by each VNF for network slice services over a continuous period of time is input into the node resource prediction model of the physical node where the VNF is located, and the resource prediction result required by the network slice at the next moment is output. The node resource prediction model includes CNN and LSTM. CNN extracts performance indicator features from performance indicators, concatenates the performance indicator features according to the time dimension to form an indicator sequence, and inputs the indicator sequence into the LSTM pair to obtain the prediction result of the resources required by the network slice at the next moment.

2. A network slice resource prediction method based on ring federated learning according to claim 1, characterized in that: When training the network slice resource prediction model, the node resource prediction model parameters trained on each node are transferred to the next node in turn as the global model parameters for training the next node, and the local VNF performance data on the next node is used to train the local model of the node.

3. The network slice resource prediction method based on ring federated learning according to claim 1, characterized in that: The process of training the network slice resource prediction model using the local VNF performance data of each node is as follows: The local VNF performance data of each node is input into the node resource prediction model of the corresponding node, and the network slice resource prediction model is trained. During training, it is assumed that the node resource prediction model parameters of the i-th client in the r-th round are In the r+1th round, the previous client j on the i-th client topology will train the node resource prediction model parameters Passed to the i-th client, client i uses it as the global model parameter of the r+1th round and The node resource prediction model of client i is trained using the local VNF performance data of client i to obtain the node resource prediction model parameters updated in round r+1.

4. The network slice resource prediction method based on ring federated learning according to claim 1, characterized in that: When training the network slice resource prediction model, the stochastic gradient descent method is used to update the parameters of the node resource prediction model.

5. The network slice resource prediction method based on ring federated learning according to claim 1, characterized in that: When training the network slice resource prediction model, the cross entropy loss is calculated for the node resource prediction model, and the cross entropy loss is used to constrain the training of the node resource prediction model.

6. A network slice resource prediction system based on ring federated learning, characterized in that: include: The data acquisition module is used to obtain the performance indicator data required by each VNF in the VNFs sequence to provide network slicing services in a continuous time; The network slice required resource prediction module is used to obtain the prediction result of the resources required by the network slice at the next moment based on the performance indicator data required by each VNF in the VNFs sequence to provide network slice services in a continuous time and the trained network slice resource prediction model. When training the network slice resource prediction model, the triplet loss is used to constrain the update direction of the node resource prediction model. The triplet loss L tl for: Among them, d represents the Euclidean distance, margin represents the threshold, and the node resource prediction model parameters of the i-th client in the r-th round are The calculated hidden layer output at time t is The hidden layer output at time t calculated by the global model on the i-th client in the r+1th round is Updated node resource prediction model The calculated hidden layer output at time t is The network slice resource prediction model is built using a circular federated learning framework. Each physical node where the VNF is located is treated as a client participating in federated learning. A node resource prediction model is built for each client. The performance indicator data required by each VNF for network slice services over a continuous period of time is input into the node resource prediction model of the physical node where the VNF is located, and the resource prediction result required by the network slice at the next moment is output. The node resource prediction model includes CNN and LSTM. CNN extracts performance indicator features from performance indicators, concatenates the performance indicator features according to the time dimension to form an indicator sequence, and inputs the indicator sequence into the LSTM pair to obtain the prediction result of the resources required by the network slice at the next moment.

7. An electronic device, characterized in that: It includes a memory and a processor, and computer instructions stored in the memory and running on the processor. When the computer instructions are run by the processor, the steps of a network slice resource prediction method based on ring federated learning as described in any one of claims 1 to 5 are completed.

8. A computer-readable storage medium, characterized in that Used to store computer instructions, which, when executed by a processor, complete the steps of a network slice resource prediction method based on ring federated learning as described in any one of claims 1 to 5.

Images