Network traffic sequence estimation method, system and storage medium based on diffusion model

Abstract

The present invention belongs to the field of network engineering and artificial intelligence technology, in particular to a method, system and storage medium for estimating network traffic sequence based on a diffusion model. The present invention proposes a method for estimating network traffic sequence based on a diffusion model, which obtains a routing matrix A of a network to be estimated and a link load sequence y at a target time point, and randomly samples an initial traffic sequence x from standard Gaussian noise. T ; Combined with the diffusion step number t, the flow sequence x t Perform reverse diffusion to obtain the flow sequence x t‑1 And perform gradient update; loop iteration until x1 is obtained, the denoising network performs denoising processing and outputs the source point flow estimation sequence x'0. The present invention can achieve high-precision network flow matrix estimation when the source point flow is missing.

Description

Technical Field

[0001] The present invention belongs to the technical field of network engineering and artificial intelligence, and in particular to a method, system and storage medium for estimating a network flow sequence based on a diffusion model. Background Art

[0002] Traffic Matrix (TM) is used to represent the traffic demand between all possible network nodes. The traffic between network nodes is often called source point traffic (OD-flows). The network status reflected by TM is of great significance in many network management issues, such as traffic engineering, anomaly detection and capacity planning. However, with the continuous expansion of the network scale, it is no longer practical to directly measure the source point traffic by collecting the transmission trajectory of each data packet. At present, a more feasible solution is to use network tomography (NT) to solve the underdetermined equations and estimate the TM at the corresponding time point by the link-level traffic (link load).

[0003] In the past decade, more and more scholars have begun to work in the field of network traffic estimation. The current mainstream TM estimation methods can be roughly divided into three categories: solutions based on premise assumptions, solutions based on mapping learning, and solutions based on generative models.

[0004] The first type of scheme makes assumptions about the distribution of TM in order to obtain the most accurate estimate possible, such as the Poisson distribution and the gravity model. This estimation scheme is extremely dependent on the established premise.

[0005] The second type of mapping learning-based solution uses deep neural networks to learn the correspondence between link load and source traffic, and then uses the trained model to directly map the link load into TM with input; currently representative ones include: introducing back propagation neural network (BPNN) to estimate the network traffic matrix and traffic estimation model based on deep belief network (DBN). This method is a fully supervised learning solution, which means that not only the traffic matrix but also the complete link load measurement is required during training, and it is necessary to ensure that the network routing and other information remain unchanged.

[0006] The third type of solution based on generative models attempts to use the learned distribution of deep generative models to restore the missing traffic matrix information in the link load. This method first samples the traffic matrix to train a generative network, and then uses the gradient of the TME objective function to search for the optimal solution in the latent space through stochastic gradient descent optimization. For example, variational autoencoders (VAE) and generative adversarial networks (GANs) are used to solve the inverse problem of TM estimation.

[0007] Although methods based on generative models can be performed with minimal assumptions and requirements, almost all deep learning methods require sufficient and complete traffic data to support model training. In addition, VAE and GAN are already unable to perform traffic matrix estimation tasks for large networks, let alone estimate more difficult network traffic sequences. Considering that there are basically missing and errors when collecting network traffic training sets, no one has yet proposed a generative network traffic sequence estimation solution that can minimize constraints, provide higher accuracy, and fit the actual situation. Summary of the invention

[0008] In order to overcome the defect in the above-mentioned prior art that the application of deep learning models in TM estimation is difficult to achieve ideal results due to the lack of source point traffic, the present invention proposes a network traffic sequence estimation method based on a diffusion model, which can achieve high-precision network traffic matrix estimation in the absence of source point traffic.

[0009] The present invention proposes a method for estimating network traffic sequence based on a diffusion model, comprising the following steps:

[0010] SA1. Obtain the routing matrix A of the network to be estimated and the link load sequence y at the target time point, and obtain the initial traffic sequence x by random sampling from standard Gaussian noise. T ; T is the maximum number of diffusion steps set;

[0011] SA2, combined with the diffusion step number t on the flow sequence x t Perform reverse diffusion to obtain the flow sequence x t-1 ;The initial value of t is T;

[0012] SA3, x t-1 Update to x t-1 +ρ t-1 ||yA×x θ (x t ,t)||2;

[0013] Among them, ||·||2 means taking the two-norm; ||yA×x θ (x t ,t)||2 means||yA×x θ (x t ,t)||2 pairs The derivative of t-1 is a constant that reflects the influence of the control gradient and is related to t; y is the link load sequence at the target time point, A is the routing matrix of the network; x θ (x t ,t) represents the training of the denoising network combined with the diffusion step number t for the flow sequence x tThe denoising result of

[0014] SA4, determine whether t-1≤1 is true; if not, update t to t-1 and return to step SA2; if yes, update x t-1 Input the denoising network and obtain the denoising sequence x output by the denoising network θ (x t-1 ,t-1) is output as the source point flow estimation sequence x'0.

[0015] Preferably, after obtaining the source point flow estimation sequence x'0 in SA4, the source point flow estimation sequence x'0 is iteratively optimized at each time point to obtain the final estimated flow sequence x final ; The iterative steps are as follows:

[0016] S51. Iterate the flow rate of each source point in x'0 using the following formula:

[0017] ;

[0018] The flow vector at time point r in x'0 after n+1 iterations is recorded as x(n+1,r), and the flow vector at time point r in x'0 after n iterations is recorded as x(n,r). represents the flow rate of the j-th source point in x(n+1,r), represents the flow of the jth source point pair in x(n,r); the initial value of n is 0, and the initial value of r is 1; represents the flow rate of the kth source point in x(n,r), a ij represents the value of the i-th row and j-th column in the routing matrix, a ik represents the value of the i-th row and k-th column in the routing matrix; y i,r represents the load of the ith link at time point r; Ny is the total number of links, and Nx is the total number of source pairs;

[0019] S52. Calculate y r The absolute value of the difference between A×x(n+1,r) and y is denoted as f(n+1,r). r Indicates the link load at time point r; determines whether f(n+1,r) is less than the judgment value, and the initial value of the judgment value is y r and the absolute value of the difference between A×x(1,r), where x(1,r) is the result of one iteration of the flow vector at time point r in x'0; if yes, the judgment value is updated to f(n+1,r), the set algebra h is updated to n+1, the initial value of h is 1, and n is updated to n+1, and then the process returns to step S51; if no, the process goes to step S53;

[0020] S53, statistically determine whether the number of iterations of x'0 after the value is updated reaches the set optimization number;

[0021] If no, update n to n+1, and then return to step S51;

[0022] If yes, then x(h,r) is used as the estimated flow sequence x final The flow vector at time point r; the current judgment value is y r and the absolute value of the difference between A×x(h,r), where x(h,r) is the flow vector at time point r after h iterations of x'0;

[0023] S54, determine whether to traverse all time points in x'0; if not, initialize the number of iterations n=0, update r to r+1, and return to step S51; if yes, output the flow sequence x final .

[0024] Preferably, the denoising network adopts a Transformer architecture consisting of M encoding modules and N decoding modules, where M and N are custom parameters; the training method of the denoising network includes the following steps:

[0025] St1, obtain the historical data of network traffic and perform normalization processing, and divide the normalized historical data into a network traffic sequence as a training sample according to the set observation time slot; obtain a preprocessing module, and the preprocessing module obtains a traffic preprocessing sequence containing all source point pairs of traffic according to the traffic observation sequence;

[0026] St2. Select a training sample x and obtain the mask mask corresponding to the training sample; input the training sample x into the preprocessing module, and the preprocessing module outputs the corresponding traffic preprocessing sequence Pre(x); calculate the traffic sequence x': x'=mask⊙x+([1]-mask)⊙Pre(x); where ⊙ represents the Hadamard product, and [1] represents a matrix of the same dimension as the mask mask;

[0027] St3, randomly sample from uniform distribution {1,2…T} to obtain diffusion step number t, t is an integer, and 1≤t≤T; T is the set maximum diffusion step number; sample random noise ε from standard Gaussian distribution, and calculate the flow sequence x' through reparameterization technique. The flow sequence after forward diffusion of diffusion step number t is taken as the noisy sequence x t ;

[0028] St4, the flow sequence x t The denoising network is input with the diffusion step number t, and the denoising network outputs the denoised sequence x0=x after denoising. θ (x t ,t); calculate the loss function Loss, and update the denoising network by gradient descent according to the loss function Loss;

[0029] St5, determine whether the denoising network has reached the convergence condition; if not, return to step St2; if yes, fix the denoising network.

[0030] Preferred: x t =α' t 1 / 2 x'+(1-α' t ) 1 / 2 ε; α' t =∏ t r=1 α r ; α r =1-β r β r ∈{β t ;1≤t≤T};α t =1-β t ; where ∏ represents continuous multiplication, α t , β t , α' t All represent constants; β t From the set {β t ;1≤t≤T};0<β1<β2<β3<…<β t <…<β T <1;α r and β r Is a constant.

[0031] Preferably, the loss function in St4 is: Loss = || mask⊙x-mask⊙x θ (x t ,t)||2;||·||2 means taking the binary norm.

[0032] Preferably, the preprocessing module includes an encoder, an intermediate network and a decoder connected in sequence; the encoder encodes the input traffic observation sequence and outputs the encoded feature data; the intermediate network is used to extract the spatiotemporal information of the encoded feature data; the decoder decodes the spatiotemporal encoded information to output a traffic preprocessing sequence containing all source points for traffic.

[0033] Preferably, the reverse diffusion in SA2 is expressed as follows:

[0034] ;

[0035] Among them, α t , β t , α' t , α' t-1 All represent constants; β t From the set {β t ;1≤t≤T}, and 0<β1<β2<β3<…<β t<…<β T <1;α t =1-β t , α' t =∏ t r=1 α r , α' t-1 =∏ t-1 r=1 α r ; α r =1-β r β r ∈{β t ;1≤t≤T};α r and β r Is a constant.

[0036] Preferably, t-1 = (1-α t-1 ) / α t-1 1 / 2 ; α t-1 =1-β t-1 , β t-1 From the set {β t ;1≤t≤T}.

[0037] The present invention also proposes a network traffic sequence estimation system based on a diffusion model, including a memory and a processor, wherein a computer program is stored in the memory, and the processor is connected to the memory, and the processor is used to execute the computer program to implement the network traffic sequence estimation method based on the diffusion model.

[0038] The present invention also proposes a storage medium storing a computer program, which is used to implement the network traffic sequence estimation method based on the diffusion model when executed. The advantages of the present invention are:

[0039] (1) The network traffic sequence estimation method based on the diffusion model proposed in this paper combines the denoising network to reversely diffuse the traffic sequence to solve the problem of underfitting of deep models caused by errors and missing training data that exist in a large number of real lives. For network traffic sequences, the spatiotemporal characteristics of the traffic sequence are fully explored. Under the condition of the same number of traffic matrices, the speed and accuracy are significantly improved compared with the previous solutions based on generative models.

[0040] (2) When the data set is incomplete, the present invention can utilize a two-stage training mechanism in which the preprocessing module and the denoising network are independent of each other to ensure the accuracy and robustness of the denoising network in estimating the actual network traffic sequence.

[0041] (3) The network traffic sequence estimation method based on the diffusion model proposed in the present invention optimizes the estimation by using the gradient approximation condition generation method and the expectation maximization algorithm, making full use of the information contained in the current link load and routing matrix, and at the same time benefiting from the diffusion model's ability to construct and fit the traffic sequence information, effectively reducing the complexity of the network traffic sequence estimation task. Compared with the existing technology, it also has a better restoration effect on the network traffic information.

[0042] (4) The present invention adopts unsupervised training and is a plug-and-play network traffic estimation solution that is only affected by the routing and topology of the current network and has the characteristics of strong adaptability and high flexibility. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 This is a schematic diagram of the structure of the preprocessing module and denoising network proposed in the present invention;

[0044] Figure 2 Flowchart of the denoising network training method;

[0045] Figure 3 It is a flow chart of the network traffic sequence estimation method based on the diffusion model;

[0046] Figure 4 Comparison of the accuracy of different models on the Abilene dataset;

[0047] Figure 5 Comparison of the accuracy of different models on the GÉANT dataset;

[0048] Figure 6 Comparison of distribution similarity of different models on the Abilene dataset;

[0049] Figure 7 Comparison of distribution similarity of different models on the GÉANT dataset. DETAILED DESCRIPTION

[0050] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0051] Reference Figure 1 ,In this implementation, a preprocessing module and a denoising network are constructed.

[0052] The preprocessing module includes an encoder, an intermediate network and a decoder connected in sequence; the encoder encodes the input traffic observation sequence x and outputs the encoded feature data; the intermediate network is used to extract the spatiotemporal information of the encoded feature data; the decoder decodes the spatiotemporal encoded information to output a traffic preprocessing sequence containing all source point pairs of traffic.

[0053] In specific implementation, the encoder and decoder can adopt a fully connected network, and the intermediate network can adopt a bidirectional recurrent neural network (Bi_RNN), etc.

[0054] The denoising network uses the Transformer architecture, which includes M encoding modules and N decoding modules, where M and N are custom parameters. The input of the denoising network is the traffic sequence with different noise levels, and the output is the traffic sequence data after noise reduction. The input of the denoising network can be recorded as the noisy sequence x t , the output is recorded as the denoised sequence x0. M>1, N>1, t is the number of diffusion steps obtained by random sampling from a uniform distribution; in the embodiment, M is set to 3, and N is set to 3.

[0055] The encoding module includes the first diffusion step embedding module, the first normalization layer, the first multi-head attention module, the first splicing network, the first forward layer and the second splicing network connected in sequence. The first forward layer consists of two layers of fully connected networks. Multiple encoding modules are connected in sequence, and the input of the first encoding module is the noisy sequence x input by the denoising network. t , in two adjacent encoding modules, the input of the latter encoding module is the output of the previous encoding module;

[0056] The input of the first diffusion step embedding module is the input of the encoding module. The first diffusion step embedding module is used to embed the diffusion step number t into the input data to form embedded data and output it; the input of the first splicing network is also connected to the input of the corresponding encoder module, and the input of the second splicing network is also connected to the output of the first splicing network; the output of the second splicing network is the output of the corresponding encoding module.

[0057] The decoding module includes the second diffusion step embedding module, the second normalization module, the second multi-head attention module, the third diffusion step embedding module, the third normalization module, the third splicing network, the third multi-head attention module, the fourth splicing network and the second forward layer connected in sequence. The second forward layer consists of two layers of fully connected networks. Multiple decoding modules are connected in sequence, and the input of the first decoding module is the noisy sequence x input by the denoising network. t ; In two adjacent decoding modules, the input of the latter decoding module is the output of the previous decoding module;

[0058] The input of the second diffusion step embedding module is the input of the decoding module, the input of the third splicing network is also connected to the input of the decoding module, the input of the third multi-head attention module is also connected to the output of the Mth encoding module, and the input of the fourth splicing network is also connected to the output of the third splicing network; the output of the second forward layer is the output of the decoding module. The output of the Nth decoding module is the output of the denoising network.

[0059] The second diffusion step embedding module is used to embed the diffusion step number t into its input data, and the third diffusion step embedding module is also used to embed the diffusion step number t into its input data.

[0060] The encoding module is used to extract the noisy sequence x input by the denoising network in combination with the diffusion step number t t The decoding module is used to combine the diffusion step number t and the feature data output by the Mth encoding module to denoise the noisy sequence x input by the denoising network. t Processing is performed to output the denoised sequence x0.

[0061] In this embodiment, the training method of the preprocessing module is as follows: input the flow observation sequence x into the preprocessing module, and the preprocessing module outputs the corresponding flow preprocessing sequence; let the source point pair of the observed flow in the flow observation sequence x be recorded as the observation source point pair, the flow of the observed source point pair in the flow observation sequence x be recorded as the real flow, and the flow of the observed source point pair in the flow preprocessing sequence output by the preprocessing module be recorded as the predicted flow, and calculate the difference between the real flow and the predicted flow on the observed source point pair as the preprocessing loss; update the preprocessing module according to the preprocessing loss until the preprocessing module reaches the set convergence condition, and the preprocessing module training is completed. The convergence condition of the preprocessing module is set as: the number of updates reaches the set iteration threshold, and the iteration threshold can be specifically taken in the interval [50,100].

[0062] Reference Figure 2 ,The training method of the denoising network includes the following steps.

[0063] St1, obtain the historical data of network traffic and perform normalization processing, and divide the normalized historical data into a network traffic sequence as a training sample according to the set observation time slot; it can be seen that the training sample is the observed dynamic traffic data that changes with time, that is, the traffic observation sequence x, and the traffic data at a single time point contains all or part of the source point traffic in the network;

[0064] St2, select the training sample x, and obtain the mask mask corresponding to the training sample; input the training sample x into the preprocessing module, and the preprocessing module outputs the corresponding traffic preprocessing sequence Pre (x); calculate the traffic sequence x';

[0065] x'=mask⊙x+([1]-mask)⊙Pre(x);

[0066] Where ⊙ represents the Hadamard product, and [1] represents a matrix of the same dimension as the mask.

[0067] The mask is a 01 matrix with one-to-one correspondence between elements and source points. The element positions with flow observations are marked with 1, and the element positions without flow observations are marked with 0.

[0068] St3, randomly sample from uniform distribution {1,2…T} to obtain diffusion step number t, t is an integer, and 1≤t≤T; T is the set maximum diffusion step number; sample random noise ε from standard Gaussian distribution, and calculate the flow sequence x' through reparameterization technique. The flow sequence after forward diffusion of diffusion step number t is taken as the noisy sequence x t ;

[0069] x t =α' t 1 / 2 x'+(1-α' t ) 1 / 2 ε;

[0070] α' t =∏ t r=1 α r ;

[0071] α t =1-β t ;

[0072] Among them, ∏ represents continuous multiplication, α t , β t , α' t All represent constants; 0<β1<β2<β3<…<β t <…<β T <1;α r =1-β r β r ∈{β t ;1≤t≤T};α r and β r Is a constant.

[0073] St4, the flow sequence x t The denoising network is input with the diffusion step number t, and the denoising network outputs the denoised sequence x0=x after denoising. θ (x t ,t); calculate the loss function Loss, and update the denoising network by gradient descent according to the loss function Loss;

[0074] Loss=||mask⊙x-mask⊙x θ (x t ,t)||2;

[0075] Among them, ||·||2 means taking the binary norm.

[0076] St5, determine whether the denoising network has reached the convergence condition; if not, return to step St2; if yes, fix the denoising network.

[0077] The present embodiment proposes a diffusion model-based network traffic sequence estimation method, which is used to predict an estimated traffic sequence representing the traffic distribution of the network to be estimated based on the traffic observation sequence of the network to be estimated and the mask.

[0078] Reference Figure 3 , the network traffic sequence estimation method based on diffusion model includes the following steps:

[0079] SA1. Obtain the routing matrix A of the network to be estimated and the link load sequence y at the target time point, and obtain the initial traffic sequence x by random sampling from standard Gaussian noise. T ;

[0080] SA2, combined with the diffusion step number t on the flow sequence x t Perform reverse diffusion to obtain the flow sequence x t-1 ;The initial value of t is T;

[0081] ;

[0082] α t =1-β t ;

[0083] α' t =∏ t r=1 α r =α1×α2×…×α r ×…×α t ;

[0084] α' t-1 =∏ t-1 r=1 α r =α1×α2×…×α r ×…×α t-1 ;

[0085] Among them, α t , β t , α' t , α' t-1 All represent constants; β t From the set {β t ;1≤t≤T}, and 0<β1<β2<β3<…<β t <…<β T <1; in the specific embodiment, β1=0.0001, βT =0.02; α r =1-β r β r ∈{β t ;1≤t≤T};α r and β r is a constant;

[0086] z t is a random variable sampled from a standard Gaussian distribution; x θ (x t ,t) represents the denoising network combined with the diffusion step number t for the flow sequence x t The denoising result of

[0087] SA3, x t-1 Update to x t-1 +ρ t-1 ||yA×x θ (x t ,t)||2;

[0088] Among them, ||·||2 means taking the two-norm; ||yA×x θ (x t ,t)||2 means||yA×x θ (x t ,t)||2 pairs The derivative of t-1 is a constant related to t that reflects the influence of the control gradient, ρ t = (1-α t ) / α t 1 / 2 , ρ t-1 = (1-α t-1 ) / α t-1 1 / 2 ; y is the link load sequence at the target time point, A is the routing matrix of the network;

[0089] SA4, determine whether t-1≤1 is true; if not, update t to t-1 and return to step SA2; if yes, update x t-1 Input the denoising network and obtain the denoising sequence x output by the denoising network θ (x1,1) is output as the source point flow estimation sequence x'0, and step SA5 is executed;

[0090] SA5, taking expectation maximization as the goal, iteratively optimize the flow data at each time point of the source flow estimation sequence x'0 to obtain the final estimated flow sequence x final ;

[0091] The iterative optimization process for x'0 includes the following steps:

[0092] S51. Let the flow vector at time point r in x'0 after 1 iteration be recorded as x(1,r), the flow vector at time point r in x'0 after n+1 iterations be recorded as x(n+1,r), and the flow vector at time point r in x'0 after n iterations be recorded as x(n,r). represents the flow rate of the j-th source point in x(n+1,r), represents the flow of the jth source point pair in x(n,r); the initial value of n is 0; x(0,r) is the flow vector at time point r in x'0;

[0093] ;

[0094] in, represents the flow rate of the kth source point in x(n,r), a ij represents the value of the i-th row and j-th column in the routing matrix, a ik represents the value of the i-th row and k-th column in the routing matrix; y i,r represents the load of the i-th link at time point r; Ny is the total number of links,

[0095] Nx is the total number of source point pairs;

[0096] S52. Calculate y r The absolute value of the difference between y and A×x(n+1,r) is recorded as f(n+1,r). Determine whether f(n+1,r) is less than the judgment value; r Indicates the link load at time point r in link load sequence y; the initial value of the judgment value is y r The absolute value of the difference between A×x(1,r);

[0097] If yes, update the judgment value to f(n+1,r), update the set algebra h to n+1, the initial value of h is 1, and update n to n+1, then return to step S51; if no, execute step S53;

[0098] S53, statistically determine whether the number of iterations of x'0 after the value is updated reaches the set optimization number;

[0099] If no, update n to n+1, and then return to step S51;

[0100] If yes, let x(h,r) be the estimated flow sequence x final The flow vector at time point r;

[0101] S54, determine whether to traverse all time points in x'0, that is, the set R; if not, initialize the number of iterations n=0, update r to r+1, and return to step S51; if yes, output the traffic sequence xfinal .

[0102] The above-mentioned diffusion model-based network traffic sequence estimation method including SA1-SA5 (hereinafter referred to as the method of the present invention) is verified in conjunction with specific embodiments below.

[0103] In this embodiment, in order to verify the performance of the method of the present invention, the method of the present invention is compared with four other existing deep learning models under various unknown traffic ratios in combination with different data sets.

[0104] In this embodiment, two real-world datasets, Abilene and GÉANT, are used. The network measured by the Abilene dataset contains 12 routes and 54 directed links. This dataset collects all OD-flows in the network at intervals of 5 minutes from March to September 2004. There are 23 routes and 120 directed links in the GÉANT network. This dataset was collected at intervals of 15 minutes from January to April 2004.

[0105] The four existing models selected in this embodiment are the variational autoencoder (VAE) estimation model VAE-TME, the generative adversarial network (WGAN) estimation model WGAN-TME, the back propagation neural network (BPNN) estimation model BPTME and its optimized MNETME.

[0106] VAE and WGAN first learn the true distribution of TM (traffic matrix), and then use gradient descent to find a solution that satisfies both the training set distribution and the Tomography equation. BPTME and MNETME train their own deep models and directly output TM estimates that meet the conditions from the Link-loads sequence.

[0107] In this embodiment, the source traffic in the data set is known. The method for verifying different comparison models for the data set is: first, the data set is divided into a pre-training set and a test set; then, part of the data in the pre-training set is replaced with unknown traffic in different proportions, and the pre-training set after replacement is used as the training set; each model is trained in combination with the training set, and then the model performance is tested on the test set. For the method of the present invention, the pre-processing module and the denoising network are first trained on the training set, and then steps SA1-SA5 are performed on the test set to obtain the final estimated traffic sequence x final .

[0108] In this embodiment, the normalized mean absolute error (NMAE) and the maximum mean difference (MMD) are used as filling accuracy and distribution evaluation indicators of each comparison model and the method of the present invention.

[0109] Figure 4The figure shows the normalized mean absolute errors achieved by the comparison models and the method of the present invention when tested on the test set by using VAE, WGAN, BPTME, MNETME and the method of the present invention in combination with the Abilene dataset for learning; Figure 5 The figure shows the normalized mean absolute errors achieved by the comparison models and the method of the present invention when tested on the test set by using VAE, WGAN, BPTME, MNETME and the method of the present invention in combination with the dataset GÉANT for learning.

[0110] Figure 6 The figure shows the maximum mean differences achieved by each comparison model and the method of the present invention when tested on the test set by using VAE, WGAN, BPTME, MNETME and the method of the present invention in combination with the data set Abilene for learning; Figure 7 The figure shows the maximum mean differences achieved by the comparison models and the method of the present invention when tested on the test set by using VAE, WGAN, BPTME, MNETME and the method of the present invention in combination with the dataset GÉANT for learning.

[0111] Figures 4 to 7 In the training known ratio, the ratio of known source point traffic in the traffic sequence is the ratio of known source point traffic in the prepared training set to unknown source point traffic. That is, the traffic sequence in the prepared training set is not missing, and part of the data of the traffic sequence in the prepared training set is replaced with unknown to obtain the training set; the training known ratio is the number of source point pairs in the training set where all traffic sequences retain source point traffic data / the number of source point pairs of all traffic sequences in the training set.

[0112] Combination Figure 4 , Figure 5 It can be seen that the accuracy of the comparison model and the method of the present invention is inversely proportional to the unknown ratio, especially on the high-dimensional dataset GÉANT. The accuracy of the method of the present invention is better than other deep models in any unknown ratio of any dataset. And the higher the unknown ratio, the greater the difference in accuracy between the comparison model and the method of the present invention, which shows that the method of the present invention shows stronger robustness. Figure 6 , Figure 7 It can be seen that the patented method also achieves a lower MMD score, which means that the proposed network traffic sequence estimation method does not compromise on distribution while achieving high accuracy.

[0113] Of course, it is obvious to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, but also includes the same or similar structures that can be implemented in other specific forms without departing from the spirit or essential features of the present invention. Therefore, no matter from which point of view, the embodiments should be regarded as exemplary and non-limiting, and the scope of the present invention is defined by the appended claims rather than the above description, and it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims are included in the present invention. Any reference numerals in the claims should not be regarded as limiting the claims involved.

[0114] In addition, it should be understood that although the present specification is described according to implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation modes that can be understood by those skilled in the art.

[0115] The techniques, shapes, and structural parts not described in detail in the present invention are all well-known techniques.

Claims

1. A network traffic sequence estimation method based on a diffusion model, characterized in that: The following steps are involved: SA1. Obtain the routing matrix A of the network to be estimated and the link load sequence y at the target time point, and obtain the initial traffic sequence x by random sampling from standard Gaussian noise. T ; T is the maximum number of diffusion steps set; SA2, combined with the diffusion step number t on the flow sequence x t Perform reverse diffusion to obtain the flow sequence x t-1 ;The initial value of t is T; SA3, x t-1 Update to x t-1 +ρ t-1 ||yA×x θ (x t ,t)||2; Among them, ||·||2 means taking the two-norm; ||yA×x θ (x t ,t)||2 means||yA×x θ (x t ,t)||2 pairs The derivative of t-1 is a constant that reflects the influence of the control gradient and is related to t; y is the link load sequence at the target time point, A is the routing matrix of the network; x θ (x t ,t) represents the training of the denoising network combined with the diffusion step number t for the flow sequence x t The denoising result of SA4, determine whether t-1≤1 is true; if not, update t to t-1 and return to step SA2; if yes, update x t-1 Input the denoising network and obtain the denoised sequence x output by the denoising network θ (x t-1 ,t-1) is output as the source point flow estimation sequence x'0.

2. The network traffic sequence estimation method based on the diffusion model according to claim 1, characterized in that: After obtaining the source point flow estimation sequence x'0 in SA4, the source point flow estimation sequence x'0 is iteratively optimized at each time point to obtain the final estimated flow sequence x final ; The iterative steps are as follows: S51. Iterate the flow rate of each source point in x'0 using the following formula: ； Among them, the flow vector at time point r in x'0 after n+1 iterations is recorded as x(n+1,r), and the flow vector at time point r in x'0 after n iterations is recorded as x(n,r). represents the flow rate of the j-th source point in x(n+1,r), represents the flow of the jth source point pair in x(n,r); the initial value of n is 0, and the initial value of r is 1; represents the flow rate of the kth source point in x(n,r), a ij represents the value of the i-th row and j-th column in the routing matrix, a ik represents the value of the i-th row and k-th column in the routing matrix; y i,r represents the load of the ith link at time point r; Ny is the total number of links, and Nx is the total number of source pairs; S52. Calculate y r The absolute value of the difference between A×x(n+1,r) and y is denoted as f(n+1,r). r Indicates the link load at time point r; determines whether f(n+1,r) is less than the judgment value, and the initial value of the judgment value is y r and the absolute value of the difference between A×x(1,r), where x(1,r) is the result of one iteration of the flow vector at time point r in x'0; if yes, the judgment value is updated to f(n+1,r), the set algebra h is updated to n+1, the initial value of h is 1, and n is updated to n+1, and then the process returns to step S51; if no, the process executes step S53; S53, statistically determine whether the number of iterations of x'0 after the value is updated reaches the set optimization number; If no, update n to n+1, and then return to step S51; If yes, then x(h,r) is used as the estimated flow sequence x final The flow vector at time point r; the current judgment value is y r and the absolute value of the difference between A×x(h,r), where x(h,r) is the flow vector at time point r after h iterations of x'0; S54, determine whether to traverse all time points in x'0; if not, initialize the number of iterations n=0, update r to r+1, and return to step S51; if yes, output the flow sequence x final .

3. The network traffic sequence estimation method based on diffusion model according to claim 1, characterized in that: The denoising network uses a Transformer architecture consisting of M encoding modules and N decoding modules, where M and N are both custom parameters. The training method of the denoising network includes the following steps: St1, obtain the historical data of network traffic and perform normalization processing, and divide the normalized historical data into a network traffic sequence as a training sample according to the set observation time slot; obtain a preprocessing module, and the preprocessing module obtains a traffic preprocessing sequence containing all source point pairs of traffic according to the traffic observation sequence; St2. Select a training sample x and obtain the mask mask corresponding to the training sample; input the training sample x into the preprocessing module, and the preprocessing module outputs the corresponding traffic preprocessing sequence Pre(x); calculate the traffic sequence x': x'=mask⊙x+([1]-mask)⊙Pre(x); where ⊙ represents the Hadamard product, and [1] represents a matrix of the same dimension as the mask mask; St3, randomly sample from uniform distribution {1,2…T} to obtain diffusion step number t, t is an integer, and 1≤t≤T; T is the set maximum diffusion step number; sample random noise ε from standard Gaussian distribution, and calculate the flow sequence x' through reparameterization technique. The flow sequence after forward diffusion of diffusion step number t is taken as the noisy sequence x t ; St4, the flow sequence x t The denoising network is input with the diffusion step number t, and the denoising network outputs the denoised sequence x0=x after denoising. θ (x t ,t); calculate the loss function Loss, and update the denoising network by gradient descent according to the loss function Loss; St5, determine whether the denoising network has reached the convergence condition; if not, return to step St2; if yes, fix the denoising network.

4. The method for estimating network traffic sequence based on diffusion model according to claim 3, characterized in that: x t =α' t 1 / 2 x'+(1-α' t ) 1 / 2 ε; α' t =∏ t r=1 α r ; α r =1-β r β r ∈{β t ;1≤t≤T};α t =1-β t ; where ∏ represents continuous multiplication, α t , β t , α' t All represent constants; β t From the set {β t ;1≤t≤T};0<β1<β2<β3<…<β t <…<β T <1;α r and β r Is a constant.

5. The network traffic sequence estimation method based on diffusion model as claimed in claim 3, characterized in that: The loss function in St4 is: Loss = || mask⊙x-mask⊙x θ (x t ,t)||2;||·||2 means taking the binary norm.

6. The network traffic sequence estimation method based on diffusion model as claimed in claim 3, characterized in that: The preprocessing module includes an encoder, an intermediate network and a decoder connected in sequence; the encoder encodes the input flow observation sequence and outputs the encoded feature data; The intermediate network is used to extract the spatiotemporal information of the encoded feature data; The decoder decodes the spatiotemporal encoded information to output a traffic preprocessing sequence containing all source point pairs of traffic.

7. The network traffic sequence estimation method based on diffusion model according to claim 1, characterized in that: The reverse diffusion in SA2 is expressed as follows: ； Among them, α t , β t , α' t , α' t-1 All represent constants; β t From the set {β t ;1≤t≤T}, and 0<β1<β2<β3<…<β t <…<β T <1;α t =1-β t , α' t =∏ t r=1 α r , α' t-1 =∏ t-1 r=1 α r ; α r =1-β r β r ∈{β t ;1≤t≤T};α r and β r Is a constant.

8. The network traffic sequence estimation method based on diffusion model according to claim 7, characterized in that: ρ t-1 =(1-α t-1 ) / α t-1 1 / 2 ; α t-1 =1-β t-1 , β t-1 From the set {β t ;1≤t≤T}.

9. A network traffic sequence estimation system based on a diffusion model, characterized in that: It comprises a memory and a processor, wherein a computer program is stored in the memory, the processor is connected to the memory, and the processor is used to execute the computer program to implement the network traffic sequence estimation method based on the diffusion model as described in any one of claims 1 to 8.

10. A storage medium, characterized in that: A computer program is stored, and when the computer program is executed, it is used to implement the network traffic sequence estimation method based on the diffusion model as described in any one of claims 1-8.

Images