5G traffic local area network slicing method based on priority and CNN-LSTM

Abstract

A 5G traffic local area network slicing method based on priorities and CNN-LSTM comprises the following steps: firstly, carrying out priority division on different service nodes of an access network and carrying out priority division on various communication messages, and further establishing a mapping mechanism from service priorities to network slices; and SLA (Service Level Agreement) setting is carried out on the network slices bearing different services. A dynamic service traffic demand is used as an input signal to be input into a CNN model, the CNN model has strong feature learning ability, and the signal is output to an LSTM model after multiple operations of a CNN convolutional layer and a pooling layer. The long-short-term memory neural network can perform feature extraction in combination with forward information, so that the defect of CNN feature extraction is overcome, global and local information is fully utilized, and the accuracy of service traffic prediction is greatly improved. And finally, maximizing the service level agreement (SLA) satisfaction rate of each slice according to the prediction result of the service demand, and realizing the optimal allocation of the wireless resources of the slices.

Description

Technical Field

[0001] The present invention belongs to the field of communication technology in intelligent transportation, and in particular to a 5G transportation local area network network slicing method based on priority and CNN-LSTM. Background Art

[0002] At present, large and medium-sized cities in China are actively deploying new infrastructure key areas such as Internet of Vehicles and Vehicle-Road Collaboration, and have made technical breakthroughs in fully closed scenarios (ports, airports), semi-closed scenarios (highways, trunk logistics, etc.), and open scenarios (driverless taxis, etc.). However, the relevant achievements have not yet been widely applied in urban traffic management scenarios. The key is that a large number of traditional traffic control equipment still lack reliable and trusted interconnection means. The 5G network has achieved full coverage in urban areas in my country, providing a solid technical condition and application foundation for breaking through the problem of trusted interconnection of transportation facilities. By using the existing 5G air interface technology physical link, the information exchange between roadside traffic information equipment (signal lights, signal control machines, radars, cameras, ground sensing coils, etc.) and moving vehicle clusters and traffic management centers can be realized, and then control to form a relatively closed industry-specific network, forming a 5G traffic LAN. The 5G traffic LAN adopts a 5G hybrid private network solution. The core network user plane network element UPF is privately deployed for the transportation industry. Wireless base stations and core network control plane network elements are flexibly deployed according to the needs of the transportation industry, providing the transportation industry with a partially physical exclusive 5G dedicated network. In this mode, the business data of users in the transportation industry network is locally unloaded, and through functional customization and optimization, the impact of public network failures on user production and business can be reduced, ensuring the reliable interconnection of transportation facilities and the secure transmission of traffic management information.

[0003] The network slicing technology in the 5G mobile communication network is a multiple isolated virtual networks created on the physical network as the infrastructure. While providing customized capabilities to meet industry requirements, it also provides certain isolation capabilities and operates independently. It is one of the key technologies to ensure the trusted interconnection of transportation facilities in the 5G transportation LAN. The 5G network can provide different slices for different industries, such as subway companies, transportation, government, banks, etc. According to the resource occupation characteristics of different industries on the 5G network, the current slice types can be divided into three categories: QoS priority guarantee slices, RB resource reservation slices, and carrier isolation. The wireless access network slices need to meet dynamic, real-time, and diverse slice communication requests. The current research direction is to use machine learning methods to achieve real-time dynamic management of slice wireless resources. However, the 5G communication requirements are very different in the public network scenario and the transportation LAN scenario. The transportation LAN communication wireless environment is more complex. Therefore, it is necessary to carry out targeted slice design for the communication network architecture of the transportation LAN, considering both the dynamic and efficient allocation of resources and the guarantee of service priority and the isolation between services, so as to achieve efficient slice allocation of limited spectrum resources, meet the needs of multiple communication services, and ensure service priority and isolation. Summary of the invention

[0004] In order to ensure the trusted interconnection of various types of transportation facilities in the 5G transportation LAN and solve the problem of efficient slicing allocation of spectrum resources while ensuring service priority and isolation, the present invention proposes a network slicing method that combines service priority with a deep learning algorithm in the 5G transportation LAN.

[0005] A 5G traffic LAN network slicing method based on priority and CNN-LSTM includes the following steps:

[0006] Step (1) prioritizing device nodes connected to the 5G traffic local area network and prioritizing various communication services of the device nodes;

[0007] Step (2) mapping the service priority to the network slice, and after mapping, setting the service level agreement (SLA) for the network slices carrying different services;

[0008] Step (3) uses the dynamic business traffic demand as an input signal, constructs a convolutional neural network model, and randomly initializes the training parameters in the CNN;

[0009] Step (4) uses batch normalization layer in convolutional layer to increase model stability and adopts tanh activation function;

[0010] Step (5) inputs the input signal to the convolution layer to extract the features of the sample data, and the subsampling layer samples the features;

[0011] Step (6) Enter the pooling layer to select extracted features and filter information, and downsample to reduce the amount of information transmitted downward to avoid overfitting;

[0012] Step (7) repeats steps (3) to (6) multiple times to train the model and calculate the loss function of the model and back-propagate the loss function value, and calculates and updates the training parameters in the CNN model until the model training termination condition is met;

[0013] Step (8) inputs the feature extraction data of the last pooling of the pooling layer into the LSTM model;

[0014] Step (9) The LSTM model uses the Tanh activation function to enhance the nonlinear fitting capability, and a Dropout layer is used in each layer to discard the output results of some neurons;

[0015] Step (10) outputs the LSTM result to the output layer, and uses the Softmax function to determine the service traffic signal status;

[0016] Step (11) finally flattens the output of the first unit and the output of the last unit in the last layer of LSTM and inputs them into the fully connected layer;

[0017] Step (12) uses a dropout layer and a Relu activation function in the fully connected layer to convert the finally extracted features into a feature vector through the fully connected layer and input it into the fully connected network to perform the classification task and output the prediction result;

[0018] Step (13) Repeat the above steps (3) to (13) multiple times to train the model;

[0019] Step (14) maximizes the service level agreement (SLA) satisfaction rate of each slice based on the predicted business demand, and achieves optimal allocation of slice wireless resources.

[0020] First, the different service nodes accessing the network are prioritized and various types of communication messages are prioritized, and then a mapping mechanism from service priority to network slices is established. Then, according to the dynamically changing network slice service requests, the service level agreement SLA satisfaction rate of each slice is maximized, thereby achieving the optimal allocation of slice wireless resources. Among them, due to user switching caused by the high-speed movement of vehicle nodes in the traffic LAN, the network slice service request volume and priority also change in real time. Therefore, the current slice wireless resource allocation is determined by the current service type and request volume. There is a certain lag, and business demand prediction is required in advance. For this reason, the present invention proposes a method for predicting business data using a convolutional neural network CNN combined with a long short-term memory network LSTM. Convolutional neural networks have powerful feature learning capabilities, but the extraction of features by a single convolutional neural network model has the problem of insufficient extraction of some features, and CNN is mainly used for the extraction and utilization of local information, and lacks global information. The long short-term memory neural network LSTM is a special recurrent neural network that is highly sensitive to time signals. It mainly acts on the processing of forward information, can capture the long-term dependency of information data well, and is suitable for business demand prediction in traffic LANs.

[0021] Compared with the prior art, the present invention has the following beneficial effects:

[0022] 1. The present invention adopts a priority-based network slicing method. In the 5G traffic LAN scenario with a more complex wireless environment, it can achieve efficient slicing allocation of limited spectrum resources while meeting multiple communication service requirements and ensuring service priority and isolation.

[0023] 2. The present invention proposes to combine the CNN model with the LSTM model to predict business traffic. The high-speed movement of vehicle nodes in the traffic LAN will cause the network slice business request volume and priority to change in real time. Forecasting business demand in advance can eliminate the lag caused by determining the current slice wireless resource allocation based on the current business type and request volume.

[0024] 3. The CNN of the present invention is mainly used for extracting and utilizing local information, but lacks global information; the inherent characteristics of LSTM lead to its failure to consider local features. The present invention proposes a CNN-LSTM model set, which complements the strengths of the two models and can fully solve the problem of extracting and utilizing global and local information, greatly improve the feature extraction capability of a single model, and effectively reduce the number of iterations required for the operation of the original model and improve the prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 Flow chart of the 5G traffic LAN network slicing method based on priority and CNN-LSTM in the present invention;

[0026] Figure 2 The present invention is based on the structural model diagram of the CNN algorithm;

[0027] Figure 3 The present invention is based on the flow chart of LSTM algorithm. DETAILED DESCRIPTION

[0028] like Figure 1 As shown, a 5G traffic LAN network slicing method based on priority and CNN-LSTM includes the following steps:

[0029] Step (1) prioritizing device nodes connected to the 5G traffic local area network and prioritizing various communication services of the device nodes;

[0030] Step (2) mapping the service priority to the network slice, and after mapping, setting the service level agreement (SLA) for the network slices carrying different services;

[0031] Step (3) uses the dynamic business traffic demand as an input signal, constructs a convolutional neural network model, and randomly initializes the training parameters in the CNN;

[0032] Step (4) uses batch normalization layer in convolutional layer to increase model stability and adopts tanh activation function;

[0033] Step (5) inputs the input signal to the convolution layer to extract the features of the sample data, and the subsampling layer samples the features;

[0034] Step (6) Enter the pooling layer to select extracted features and filter information, and downsample to reduce the amount of information transmitted downward to avoid overfitting;

[0035] Step (7) repeats steps (3) to (6) multiple times to train the model and calculate the loss function of the model and back-propagate the loss function value, and calculates and updates the training parameters in the CNN model until the model training termination condition is met;

[0036] Step (8) inputs the feature extraction data of the last pooling of the pooling layer into the LSTM model;

[0037] Step (9) The LSTM model uses the Tanh activation function to enhance the nonlinear fitting capability, and a Dropout layer is used in each layer to discard the output results of some neurons;

[0038] Step (10) outputs the LSTM result to the output layer, and uses the Softmax function to determine the service traffic signal status;

[0039] Step (11) finally flattens the output of the first unit and the output of the last unit in the last layer of LSTM and inputs them into the fully connected layer;

[0040] Step (12) uses a dropout layer and a Relu activation function in the fully connected layer to convert the finally extracted features into a feature vector through the fully connected layer and input it into the fully connected network to perform the classification task and output the prediction result;

[0041] Step (13) Repeat the above steps (3) to (13) multiple times to train the model;

[0042] Step (14) maximizes the service level agreement (SLA) satisfaction rate of each slice based on the predicted business demand, and achieves optimal allocation of slice wireless resources.

[0043] like Figure 2 As shown, the present invention constructs a CNN model, comprising the following steps:

[0044] Step S1-1: Dynamic service traffic demand d = {d1, d2, ..., d m},m∈{1,2,…,m}, It is input into the CNN model as an input signal and enters the convolutional layer for normalization and feature extraction;

[0045] Step S1-2: Randomly initialize the training parameters in CNN, use batch normalization layer in convolution layer to normalize data to enhance generalization ability and use tanh activation function

[0046] Step S1-3: Input the input signal to the convolution layer, and use the initial input data of 32×32 with a depth of 3 in the input layer; after entering the convolution layer, use a convolution kernel of size 7×7×3 on the input data and slide 1 unit in turn according to the corresponding formula to extract features of the sample data. At the same time, the subsampling layer also samples the features. After feature extraction, draw a feature extraction graph according to the feature values ​​obtained by convolution;

[0047] Step S1-4: In order to obtain richer feature data, convolution kernels of multiple scales are used to extract features from the input data; for the first time, 6 convolution kernels of scale 7×7×3, step size 1 and border padding 2 are used to perform multiple feature extractions to obtain 6 different feature maps. On this basis, multiple convolution kernels of different scales are used to perform multiple feature extractions.

[0048] Step S1-5: The feature data enters the pooling layer to start the pooling operation, and the final feature map obtained by multiple convolutions in step S1-4 is compressed to filter useful information from a large amount of data, and downsamples to reduce the amount of information transmitted downward to avoid overfitting;

[0049] Step S1-6: Repeat steps S1-3 to S1-5 multiple times to perform model training and calculate the loss function of the model and back-propagate the loss function value, calculate and update the training parameters in the CNN model until the model training termination conditions are met.

[0050] like Figure 3 As shown, the LSTM algorithm process of the present invention is specifically as follows:

[0051] Step S2-1: Input the feature extraction data of the last pooling of the pooling layer into the LSTM model, and generate the forward propagation layer h(t)=f(Wx t +Uh t-1 );

[0052] Where f(·) is the propagation formula; W is x t and h t-1 The connection weight matrix between t-1 With h t The connection weight matrix between t Output value for LSTM;

[0053] Step S2-2: Add Tanh activation function to LSTM model Enhance the nonlinear fitting ability, and use the Dropout layer in each layer to discard the output results of some neurons;

[0054] Step S2-3: Output the LSTM result to the output layer, and use the Softmax function to determine the service traffic signal status;

[0055] Step S2-4: Flatten the output of the first unit and the output of the last unit of the last layer of LSTM and input them into the fully connected layer;

[0056] Step S2-5: Use Dropout layer regularization and Relu activation function Relu=max(0,x) in the fully connected layer, convert the finally extracted feature data into corresponding feature vectors through the fully connected layer, input them into the fully connected network, perform classification tasks, and output prediction results;

[0057] Step S2-6: Repeat the above steps S2-1 to S2-5 multiple times to train the model.

[0058] The 5G traffic LAN network slicing method based on priority and CNN-LSTM includes the following steps:

[0059] Step S3-1: Prioritize various nodes connected to the 5G traffic LAN, and prioritize various communication services of various nodes;

[0060] Step S3-2: Mapping service priorities to network slices, and setting service level agreements (SLAs) for network slices carrying different services;

[0061] Step S3-3: Dynamic service traffic demand is expressed as d = {d1, d2, ..., d m},m∈{1,2,…,M}, Q u A set of data packets transmitted by the base station, including characteristic information of various services carried by the slice;

[0062] Step S3-4: construct a CNN convolutional neural network model, input the dynamic service traffic demand as an input signal into the CNN model, and enter the convolution layer for normalization and feature extraction;

[0063] Step S3-5: The feature data extracted by the convolution layer is input into the pooling layer for pooling and compression of features to screen useful feature information;

[0064] Step S3-6: Repeat steps S4 and S5 multiple times to calculate and update the training parameters in the CNN model until the termination condition is met;

[0065] Step S3-7: Input the last pooling result into the LSTM model, flatten and flatten the output of the first unit and the output of the last unit of the last layer of the LSTM, and input them into the fully connected layer;

[0066] Step S3-8: Convert feature data in the fully connected layer and perform classification tasks, and output the business traffic prediction results;

[0067] Step S3-9: According to the predicted business demand, maximize the service level agreement SLA satisfaction rate SSR of each slice to achieve optimal allocation of slice wireless resources;

[0068] in, The value is 0 or 1. When the base station and slice m user u can transmit data packet q within the communication transmission time u The transmission is successful under the conditions of meeting the service level agreement. The value is 1, otherwise it is 0.

[0069] Principle of the present invention:

[0070] The 5G traffic LAN network slicing method based on priority and CNN-LSTM of the present invention first prioritizes different service nodes accessing the network and prioritizes various types of communication messages, and then establishes a mapping mechanism from service priority to network slices, and sets a service level agreement SLA for network slices carrying different services. The dynamic service traffic demand is input as an input signal into the CNN model, which has a powerful feature learning ability. The signal is output to the LSTM model after multiple operations of the CNN convolution layer and the pooling layer. The long short-term memory neural network can extract features in combination with forward information, improve the deficiencies of CNN feature extraction, make full use of global and local information, and greatly improve the accuracy of service traffic prediction. Finally, according to the predicted results of business demand, the service level agreement SLA satisfaction rate of each slice is maximized to achieve the optimal allocation of slice wireless resources.

Claims

1. A 5G traffic LAN network slicing method based on priority and CNN-LSTM, characterized in that The steps include: Step (1) prioritizing device nodes connected to the 5G traffic local area network and prioritizing various communication services of the device nodes; Step (2) mapping the service priority to the network slice, and after mapping, setting the service level agreement (SLA) for the network slices carrying different services; Step (3) uses the dynamic business traffic demand as an input signal, constructs a convolutional neural network model, and randomly initializes the training parameters in the CNN; Step (4) uses batch normalization layer in convolutional layer to increase model stability and adopts tanh activation function; Step (5) inputs the input signal to the convolution layer to extract the features of the sample data, and the subsampling layer samples the features; Step (6) Enter the pooling layer to select extracted features and filter information, and downsample to reduce the amount of information transmitted downward to avoid overfitting; Step (7) repeats steps (3) to (6) multiple times to train the model and calculate the loss function of the model and back-propagate the loss function value, and calculates and updates the training parameters in the CNN model until the model training termination condition is met; Step (8) inputs the feature extraction data of the last pooling of the pooling layer into the LSTM model; Step (9) The LSTM model uses the Tanh activation function to enhance the nonlinear fitting capability, and a Dropout layer is used in each layer to discard the output results of some neurons; Step (10) outputs the LSTM result to the output layer, and uses the Softmax function to determine the service traffic signal status; Step (11) finally flattens the output of the first unit and the output of the last unit in the last layer of LSTM and inputs them into the fully connected layer; Step (12) uses a dropout layer and a Relu activation function in the fully connected layer to convert the finally extracted features into a feature vector through the fully connected layer and input it into the fully connected network to perform the classification task and output the prediction result; Step (13) Repeat the above steps (3) to (13) multiple times to train the model; Step (14) maximizes the service level agreement (SLA) satisfaction rate of each slice based on the predicted business demand, and achieves optimal allocation of slice wireless resources.

2. The 5G traffic LAN network slicing method based on priority and CNN-LSTM according to claim 1 is characterized in that Building a CNN model includes the following steps: Step S1-1: Dynamic service traffic demand d = {d1, d2, ..., d m },m∈{1,2,…,M}, It is input into the CNN model as an input signal and enters the convolutional layer for normalization and feature extraction; Step S1-2: Randomly initialize the training parameters in CNN, use batch normalization layer in convolution layer to normalize data to enhance generalization ability and use tanh activation function Step S1-3: Input the input signal to the convolution layer, and use the initial input data of 32×32 with a depth of 3 in the input layer; after entering the convolution layer, use a convolution kernel of size 7×7×3 on the input data and slide 1 unit in turn according to the corresponding formula to extract features of the sample data. At the same time, the subsampling layer also samples the features. After feature extraction, draw a feature extraction graph according to the feature values ​​obtained by convolution; Step S1-4: In order to obtain richer feature data, convolution kernels of multiple scales are used to extract features from the input data; The first time, 6 convolution kernels with a scale of 7×7×3, a step size of 1 and a border padding of 2 were used to perform multiple feature extractions, and 6 different feature maps were obtained. On this basis, multiple convolution kernels of different scales were used to perform multiple feature extractions. Step S1-5: The feature data enters the pooling layer to start the pooling operation, and the final feature map obtained by multiple convolutions in step S1-4 is compressed to filter useful information from a large amount of data, and downsamples to reduce the amount of information transmitted downward to avoid overfitting; Step S1-6: Repeat steps S1-3 to S1-5 multiple times to perform model training and calculate the loss function of the model and back-propagate the loss function value, calculate and update the training parameters in the CNN model until the model training termination conditions are met.

3. The 5G traffic LAN network slicing method based on priority and CNN-LSTM according to claim 2 is characterized in that The specific process of LSTM algorithm is as follows: Step S2-1: Input the feature extraction data of the last pooling of the pooling layer into the LSTM model, and generate the forward propagation layer h(t)=f(Wx t +Uh t-1 ); Where f(·) is the propagation formula; W is x t and h t-1 The connection weight matrix between t-1 With h t The connection weight matrix between t Output value for LSTM; Step S2-2: Add Tanh activation function to LSTM model Enhance the nonlinear fitting ability, and use the Dropout layer in each layer to discard the output results of some neurons; Step S2-3: Output the LSTM result to the output layer, and use the Softmax function to determine the service traffic signal status; Step S2-4: Flatten the output of the first unit and the output of the last unit of the last layer of LSTM and input them into the fully connected layer; Step S2-5: Use Dropout layer regularization and Relu activation function Relu=max(0,x) in the fully connected layer, convert the finally extracted feature data into corresponding feature vectors through the fully connected layer, input them into the fully connected network, perform classification tasks, and output prediction results; Step S2-6: Repeat the above steps S2-1 to S2-5 multiple times to train the model.

4. The 5G traffic LAN network slicing method based on priority and CNN-LSTM according to claim 3 is characterized in that: Step S3-1: Prioritize various nodes connected to the 5G traffic LAN, and prioritize various communication services of various nodes; Step S3-2: Mapping service priorities to network slices, and setting service level agreements (SLAs) for network slices carrying different services; Step S3-3: Dynamic service traffic demand is expressed as d = {d1, d2, ..., d m },m∈{1,2,…,M}, Q u A set of data packets transmitted by the base station, including characteristic information of various services carried by the slice; Step S3-4: construct a CNN convolutional neural network model, input the dynamic service traffic demand as an input signal into the CNN model, and enter the convolution layer for normalization and feature extraction; Step S3-5: The feature data extracted by the convolution layer is input into the pooling layer for pooling and compression of features to screen useful feature information; Step S3-6: Repeat steps S4 and S5 multiple times to calculate and update the training parameters in the CNN model until the termination condition is met; Step S3-7: Input the last pooling result into the LSTM model, flatten and flatten the output of the first unit and the output of the last unit of the last layer of the LSTM, and input them into the fully connected layer; Step S3-8: Convert feature data in the fully connected layer and perform classification tasks, and output the business traffic prediction results; Step S3-9: According to the predicted business demand, maximize the service level agreement SLA satisfaction rate SSR of each slice to achieve optimal allocation of slice wireless resources; in, The value is 0 or 1. When the base station and slice m user u can transmit data packet q within the communication transmission time u The transmission is successful under the conditions of meeting the service level agreement. The value is 1, otherwise it is 0.

Images