Method for modeling user service traffic at network layer in wireless network

Abstract

Disclosed in the present invention is a method for modeling user service traffic at a network layer in a wireless network, which method is applied to the field of wireless networks. With regard to the problem of user service traffic in a wireless network environment, in the present invention, a network traffic model based on a denoising diffusion probabilistic models (DDPM) diffusion network is established, and for data packets of user service traffic which are captured by a network layer, packet sizes and packet arrival intervals of the data packets are modeled, so as to obtain a traffic model of a user service at the network layer; moreover, by means of the service traffic modeling method in the present invention, a network traffic model maintaining the multi-dimensional correlation of original service traffic while conforming to the statistical distribution of the original service traffic can be established.

Description

A modeling method for user service traffic at the network layer in wireless networks Technical Field

[0001] The present invention belongs to the field of wireless networks, and in particular relates to a user service flow modeling technology in a wireless network. Background Art

[0002] In wireless networks, network traffic modeling plays a crucial role in verifying base station communication resource scheduling and allocation strategies. With the advancement of information technology, network traffic has become increasingly diverse and complex, making it difficult to describe the true distribution of network traffic using mathematical formulas. Traditional simulation verification typically uses formulas to fit traffic data transmitted across the network. While this approach is simple and direct, consuming minimal computing resources, it also presents challenges in fitting complex traffic distributions and the multidimensional nature of traffic. It's difficult to preserve the multidimensional correlations of the original traffic data while simultaneously depicting the true traffic distribution. Furthermore, the data generated by the formula-fitting model struggles to improve the generalization capabilities of downstream task models. Therefore, establishing an efficient and accurate network traffic model is crucial to the development of wireless networks. Summary of the Invention

[0003] To solve the above technical problems, the present invention proposes a modeling method for user service traffic at the network layer in a wireless network, and establishes a network traffic model based on the DDPM diffusion model, which conforms to the statistical distribution of the original service traffic while retaining the multidimensional correlation of the original service traffic; at the same time, the MLP network structure with time step embedding is adopted to reduce the computational complexity of the model, reduce the training cost of the model and improve the generation efficiency of the model.

[0004] The technical solution adopted by the present invention is: a method for modeling user service traffic at the network layer in a wireless network, comprising:

[0005] S1. After the user service is started, use the Wireshark packet capture tool to capture the data packets of the user service flow at the network layer.

[0006] S2. Analyze the packet size and arrival time of each data packet captured by Wireshark, and construct user service traffic data based on the packet size and arrival time of each data packet.

[0007] S3. Clean the user service traffic data constructed in step S2; slice the cleaned data to construct the original data set;

[0008] S4, sample data from the original dataset, randomly sample Gaussian noise ε and time step t;

[0009] S5. Constructing a DDPM noise prediction network, wherein the DDPM noise prediction network includes a forward diffusion part and a time step embedded MLP (Multi-Layer Perceptron) network part;

[0010] The DDPM diffusion model adds noise to the input data x0 according to the randomly sampled Gaussian noise ε and the time step t. The noise addition process is:

[0011] in, is the signal coefficient. As t increases, reduce;

[0012] x t The MLP network embedded with the time step of the DDPM diffusion model and t is used to predict the noise added during forward diffusion. θ (x t ,t);

[0013] S6, according to the predicted noise ε θ (x t ,t) and the real added noise ε calculate the root mean square loss: MSELOSS=||ε-ε θ (x t ,t)|| 2

[0014] Where ε is the sampled Gaussian noise;

[0015] S7. For the calculated root mean square loss function, use the Adam optimizer to update the model parameters, set the learning rate to 0.001, and train the model until the network converges.

[0016] Step S3 also includes normalizing the packet size and packet arrival time of the cleaned data.

[0017] Beneficial effects of the present invention: The present invention provides a network layer modeling method for user service traffic in wireless networks. This method establishes a network traffic model based on the DDPM diffusion model, conforming to the statistical distribution of the original service traffic while retaining its multidimensional correlation. Furthermore, the use of a time-step embedded MLP network structure reduces the model's computational complexity, reduces model training costs, and improves model generation efficiency. The method is generally applicable to establishing user service traffic models in wireless networks. The present invention has the following advantages:

[0018] 1. A traffic model based on the DDPM (Denoising Diffusion Probabilistic Models) diffusion network was established, which not only conforms to the statistical distribution of the original business traffic but also retains the multi-dimensional correlation of the original business traffic;

[0019] 2. The MLP network structure with time step embedding is used to predict noise, which reduces the computational complexity of the model while ensuring the performance of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] FIG1 is a flow chart of the method of the present invention.

[0021] FIG2 is a noise prediction network of the DDPM diffusion model adopted by the present invention.

[0022] Figure 3 is a scatter plot of packet size and packet arrival interval of the traffic model and actual business flow.

[0023] FIG4 is a schematic diagram of the Pearson correlation coefficient between the packet size and packet arrival interval of the traffic model and the actual business flow. DETAILED DESCRIPTION

[0024] To facilitate those skilled in the art to understand the technical content of the present invention, the present invention is further explained below with reference to the accompanying drawings.

[0025] The technical solution of the present invention is:

[0026] FIG1 is a flow chart of the solution of the present invention, which specifically includes the following steps:

[0027] S1. After starting the user service, open the Wireshark packet capture tool to capture the data packets of the user service flow at the network layer. To ensure that the captured data is fully distributed, capture two hours of data;

[0028] S2. Analyze the packet size and arrival time of each data packet captured by Wireshark to construct user service traffic data. The data includes two dimensions: packet size and packet arrival interval.

[0029] S3. Further clean the user service traffic data, remove outliers, and normalize each dimension of the data to accelerate the convergence of the neural network. In addition, it is necessary to slice the data to construct a training data set. The continuous two-dimensional data is divided into (L, 2) data slices, and these data slices are randomly shuffled. L is the set data slice length, and the value of L is determined based on the traffic cycle of the user service flow. For example, the user service flow to be modeled is plotted and the traffic cycle is observed. The data packet format corresponding to a traffic cycle can be selected as L.

[0030] S4. When the current batch is in the epoch, start training the DDPM diffusion model. Sample data x from the original training dataset obtained in step S3. Randomly sample Gaussian noise ε and time step t. Divide the data into a shape of (B, L, 2) according to the batch size B. Therefore, the random Gaussian noise shape is (B, L, 2) and the time step t is (B, L, 1).

[0031] S5, add the sampled Gaussian noise ε and time step t to the data x0 to obtain x t , x t The DDPM diffusion model first adds noise to the input data x0 based on the sampled Gaussian noise ε and time step t. The noise addition process is:

[0032] in, is the signal coefficient. As t increases, As x decreases, the original data gradually tends to Gaussian noise. t and t are fed into the noise prediction network θ, and the predicted noise e = ε θ (x t ,t),ε θ (x t ,t) represents the noise prediction network θ at input x t and the predicted output at time t.

[0033] S6. Calculate the root mean square loss based on the predicted noise e and the actual added noise n. Calculate the root mean square loss function of the neural network predicted noise e and the sampled noise ε, and the formula is as follows: MSELOSS = ||ε-ε θ (x t ,t)|| 2

[0034] Where ε is the sampled Gaussian noise, ε θ (x t ,t) is the noise predicted by the model.

[0035] S7. Use the Adam optimizer to backpropagate and update the model parameters of the DDPM noise prediction network, specifically the weights and biases between neurons. Based on the calculated root mean square loss function, use the Adam optimizer to update the model parameters, setting the learning rate to 0.001, and train the model until the network converges. Specifically, if the learning rate of the neural network is set to 0.001 during training, the network is considered converged when the loss change per epoch is less than 0.001.

[0036] As shown in FIG2, the MLP network structure diagram of the time step embedding of the DDPM diffusion model of the present invention is shown. In FIG2, E represents the embedding layer, Linear represents the linear layer, each linear layer corresponds to an embedding layer, and the input of each embedding layer is the current time step t. The input of each linear layer includes the output of the previous linear layer and the output of the corresponding embedding layer, where the input of the first current layer is x. t The output of the last linear layer is the predicted noise. The present invention uses this network to predict the noise ε added at the current time step t. θ (x t ,t), for the current input data x0, sampling Gaussian noise ε and random time step t, forward diffusion is obtained:

[0037] x t The MLP network embedded with t input time step predicts the noise added during forward propagation as ε θ (x t ,t).

[0038] Model generation method: The process of generating traffic is a reverse process, that is, a denoising process. T Starting from N(0,I), we can generate a real sample by gradually removing noise. t-1 |x t The entire training sample is needed, given x t and t, we can use the model θ to estimate a sampling distribution By calculating the difference from the true distribution, we can achieve the goal of maximizing the likelihood probability. We also define the reverse process as a Markov chain, except that it is composed of a series of Gaussian distributions parameterized by neural networks: P θ (x t-1 |x t )=N(x t-1 ;μ θ (x t ,t),Σ θ (x t ,t)) z~N(0,1)

[0039] I represents an identity matrix, which is a square matrix with all diagonal elements being 1 and all off-diagonal elements being 0. t-1 ;μ θ (x t ,t),Σ θ (x t ,t)) represents x t-1Subject to the mean μ θ (x t ,t), the variance is Σ θ (x t ,t)’s Gaussian distribution; here P θ (x T )=N(x T ;0,I),N(x T ; 0,1) represents x T It obeys a Gaussian distribution with a mean of 0 and a variance of I, and P θ (x t-1 |x t ) is a parameterized Gaussian distribution, P θ (x T ) and P θ (x t-1 |x t ) is obtained by training the network μ θ (x t ,t and Σ θ (x t ,t) is given. By iterating multiple times, we can get x in order T →...→x t →x t-1 →...→x0. The diffusion model is to obtain these trained networks and use them to predict the distribution of random noise added at each step, forming the final generative model.

[0040] Figure 3 shows a scatter plot of packet size and packet arrival interval between the traffic model and actual traffic flows. The horizontal axis is the packet arrival interval, and the vertical axis is the packet size. We normalized them to between (0, 1), confirming that the data distribution generated by the traffic model is consistent with the actual traffic flows. Figure 4 shows the Pearson correlation coefficient between packet size and packet arrival interval between the traffic model and actual traffic flows. We sampled 20 times, each sampling over 6,000 consecutive data points. The Pearson correlation coefficients of the model-generated data and the original data are very consistent, demonstrating that our traffic model has learned the multi-dimensional correlations of actual traffic flows.

[0041] Those skilled in the art will appreciate that the embodiments described herein are intended to aid the reader in understanding the principles of the present invention, and it should be understood that the scope of the present invention is not limited to such specific descriptions and embodiments. Various modifications and variations are readily apparent to those skilled in the art. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention are intended to be included within the scope of the claims.

Claims

1. A method for modeling user service traffic at the network layer in a wireless network, characterized in that: include: S1. After the user service is started, use the Wireshark packet capture tool to capture the data packets of the user service flow at the network layer. S2. Analyze the packet size and arrival time of each data packet captured by Wireshark, and construct user service traffic data based on the packet size and arrival time of each data packet. S3. Clean the user service traffic data constructed in step S2; slice the cleaned data to construct the original data set; S4, sample data from the original dataset, randomly sample Gaussian noise ε and time step t; S5. Constructing a DDPM noise prediction network, wherein the DDPM noise prediction network includes a noise addition part and a noise prediction part; The noise addition part adds noise to the input data x0 according to the randomly sampled Gaussian noise ε and time step t: in, is the signal coefficient. As t increases, reduce; x t And t is input into the denoising part, and the noise added to the predicted denoising part is ε θ (x t ,t); S6, according to the predicted noise ε θ (x t ,t) and the real added noise ε calculate the root mean square loss: MSELOSS=||e-e θ (x t ,t)|| 2 Where ε is the sampled Gaussian noise; S7. For the calculated root mean square loss function, use the Adam optimizer to update the model parameters, set the learning rate to 0.001, and train the model until the network converges.

2. The method for modeling user service traffic at the network layer in a wireless network according to claim 1, characterized in that: The denoising part in step S5 specifically adopts an MLP network with time step embedding.

3. The method for modeling user service traffic at the network layer in a wireless network according to claim 2, characterized in that: Step S3 also includes normalizing the packet size and packet arrival time of the cleaned data.

4. The method for modeling user service traffic at the network layer in a wireless network according to claim 3, characterized in that: It also includes slicing the normalized two-dimensional data to obtain several data slices, randomly shuffling these data slices, and constructing a training data set.

5. The method for modeling user service traffic at the network layer in a wireless network according to claim 4, characterized in that: The length of each data slice is determined according to the traffic cycle of the user service flow.

Images