A congestion propagation path prediction method based on network information mining

Abstract

A congestion propagation path prediction method based on network information mining, modeling the data transmission network structure and its congestion state as a dynamic graph structure; screening the candidate congestion propagation path node set; at the same time, obtaining the adjacency matrix of the nodes participating in the network congestion in the candidate path set; taking the historical congestion state value of the nodes participating in the network congestion propagation as input, generating the vector representation of the congestion propagation path through a multi-layer LSTM module; taking the vector representation of the congestion propagation path and the adjacency matrix of the nodes participating in the network congestion as input, passing through a graph convolution module composed of two layers of graph convolution network and one fully connected layer, and selecting the path corresponding to the node with the highest weight in the output result as the predicted network congestion path. The application predicts the propagation path leading to the congestion state of the target node, provides effective support and suggestions for architecture adjustment, resource scheduling, fault detection, predictive maintenance and other technologies, and improves the end-to-end reliability of the network.

Description

Technical Field

[0001] The present invention belongs to the technical field of network information transmission, and relates to a network congestion path prediction method combining a graph data structure and a recurrent neural network. Background Art

[0002] With the continuous advancement of networking and informatization, today's society's production and daily life are becoming increasingly dependent on the network. Once the data transmission volume in an active computer network exceeds the network's carrying capacity, and the network nodes have limited storage and forwarding resources, and the data tasks exceed their inherent processing power and capacity, network congestion is likely to occur, leading to a decline in network transmission performance.

[0003] There are many ways to monitor network congestion. Common metrics include packet loss rate, average queue length, number of timed-out retransmitted packets, average latency and its variance, data flow size or throughput, and more. Congestion control is the primary approach to addressing network congestion. This ensures that the network can handle the data traffic required by user tasks. It is a global issue involving many factors, including hosts and routers. By analyzing and studying the global data generated by historical network congestion, we can uncover the inherent patterns underlying network congestion, derive and identify the causes of congestion and the paths through which it propagates, and refine congestion control strategies to improve network communication efficiency. Summary of the Invention

[0004] Aiming at the problem of network congestion, the present invention provides a congestion propagation path prediction method based on network information mining. It uses network node information and its historical data, combined with graph data structure and recurrent neural network, to predict the congestion propagation path that leads to target node congestion, so as to help analyze the cause of network congestion and improve congestion control strategy.

[0005] The technical solution for implementing the present invention comprises the following steps:

[0006] Step 1: Model the data transmission network and its congestion state as a dynamic graph structure {G t |G t =(V,A t )}. The independent devices participating in network data transmission are the node set V in the figure = {v i ,i=1,2,…,N v}, where N v is the total number of nodes, and the connectivity state between nodes in time slice t is the network dynamic adjacency matrix A t ,The graphs of adjacent time slices can be connected according to the connectivity of data transmission paths in the network;

[0007] Step 2: Screen the candidate congestion propagation path set of the target node where N pis the total number of candidate congested propagation paths, and the adjacency matrix A of the participating nodes in the candidate congested paths is obtained. o- ;

[0008] Step 3: Take the historical congestion status value of the node corresponding to the candidate congestion propagation path as rk,tk} as input, a vector representation H of the congestion propagation path is generated through a multi-layer LSTM module;

[0009] Step 4: Use the congestion propagation path vector representation H obtained in step 3 as the feature matrix in the graph convolutional network, and the adjacency matrix A of the participating nodes in the candidate congestion path o- As the adjacency matrix in the graph convolutional network, it is input into the graph convolution module consisting of two layers of graph convolutional network and one fully connected layer, and the path corresponding to the top d nodes in the congestion status prediction value in the output result is selected as the predicted network congestion path.

[0010] Preferably, in step 1, the dynamic adjacency matrix in the dynamic graph structure data model is defined as follows:

[0011] The total sampling time n is divided into m time slices of equal length l. The devices in the network that can independently send and receive data are regarded as nodes and are uniquely identified. For example, device i corresponds to node v i ;

[0012] If there is a directly connectable data transmission path between device i and device j, and the normalized data volume transmitted from i to j in time slice t is s ij,t , define the network dynamic association matrix The subscript t indicates the value in time slice t;

[0013] Define the network intrinsic adjacency matrix

[0014]

[0015] In summary, within a specific time slice t, the network dynamic adjacency matrix between nodes is for:

[0016] A t =S t +A o , where element A ij,t =s ij,t +a ij

[0017] Preferably, in step 2, the method for screening the candidate congested propagation path set P is as follows:

[0018] Based on the dynamic graph structure modeled in step 1, the directed connections between adjacent nodes along the time axis and between adjacent time slices represent the actual data transmission path in the network. A congestion propagation path is the set of paths formed by concatenating directly connected nodes and indirectly connected nodes leading to the target node. The set of candidate congestion propagation paths is a subset of this set.

[0019] In the present invention, all nodes on a path are connected in series to represent the path. The candidate congested propagation path set is a set of nodes that are adjacent to each other. Therefore, a candidate congested path p can be represented as:

[0020] p=(v r1 ,v r2 ,…,v rd )

[0021] The path length is the number of nodes on a single path, d.

[0022] Calculate the current node v i The in-degree sequence:

[0023]

[0024] Take the maximum in-degree max{D i,t}Corresponding time slice t D The adjacent time slice t D-1 Upper v i All adjacent nodes of the current node are connected to generate the spatiotemporal neighboring point set of the current node

[0025] Similarly, calculate its in-degree sequence

[0026] Calculate and compare the current node v i Adjacent nodes in the spatiotemporal neighborhood Pearson correlation coefficient between in-degree sequences

[0027] Take the two most relevant adjacent nodes corresponding to the largest Pearson coefficient values ​​and generate two new paths. The two adjacent nodes are used as the previous nodes of the current node in the two new paths. Let the selected adjacent nodes be the new current node;

[0028] Repeat the above process of finding the two most relevant adjacent nodes on adjacent time slices from the maximum in-degree until each path reaches the preset number of single path nodes d;

[0029] Record all new paths generated in the process to filter out all candidate paths

[0030]

[0031] From the network inherent adjacency matrix A o Remove the corresponding values ​​of the nodes that do not participate in the construction of P and obtain the adjacency matrix A of the participating nodes o- .

[0032] Preferably, in step 3, a multi-layer LSTM module is designed to obtain a vector representation of the candidate congested path. The structure of the multi-layer LSTM module is as follows:

[0033] If a node v i There are N candidate congested paths in the set p paths, where each candidate congestion propagation path corresponds to a layer of LSTM network;

[0034] For a candidate path p r =(v r1 ,v r2 ,…,v rd ), the corresponding congestion propagation time slice is T r =(t r1 ,t r2 ,…,t rd ). Then the candidate path p of this node r In the corresponding LSTM network, the input of the kth LSTM unit is p r Medium v rk In time slice t rk The node congestion state value O rk,tk , and the state h output by the k-1th LSTM unit r,k-1 ;

[0035] The state quantity h finally output by the rd LSTM unit r,rd As the path p r The eigenvalue of . Then N p The characteristic values ​​of the paths constitute the candidate path vector representation h of the node i ;

[0036] Furthermore, the candidate path vector representations of all participating network congestion nodes can be concatenated into a feature evidence N is the total number of nodes involved in network congestion.

[0037] The present invention mines network congestion propagation patterns based on the historical state information of the network and the spatiotemporal correlation between nodes, and predicts the propagation path that leads to the congestion state of the target node. The data transmission network and its congestion state are modeled as a dynamic graph structure, with the data transceiver devices in the network as nodes of the graph, the connections between nodes representing data transmission links, the adjacency matrix defined using network state data, and one time slice forming a graph. The graphs of adjacent time slices are connected according to the connectivity of network link transmission. Then, by mining the network information features in the graph structure, a set of candidate congestion propagation path nodes of the target node is generated. Finally, the historical data of the candidate congestion paths and their related nodes are input into a recurrent neural network model composed of a multi-layer LSTM and a graph convolutional network, and the predicted value of the congestion state of the candidate path nodes is finally output, thereby inferring the network congestion propagation path that leads to the congestion of the target node within the predicted time slice.

[0038] The beneficial effects of the present invention are as follows: fully considering the structural and time-varying characteristics of the network, combining congestion status data and network topology to model dynamic graph structure data with both spatial and time-varying characteristics; greatly reducing the computational complexity of the prediction model through candidate path screening, thereby improving prediction efficiency and accuracy; inferring possible congestion propagation paths through the predicted values ​​of the congestion status data of candidate path nodes, providing effective support and suggestions for network architecture adjustment, resource scheduling, fault detection, predictive maintenance and other technologies, and improving the end-to-end reliability of the network. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 is a flow chart of the method of the present invention;

[0040] Figure 2 It is a network graph structure data model and a schematic diagram of congestion propagation paths;

[0041] Figure 3 This is the architecture diagram of the path node prediction recurrent neural network. DETAILED DESCRIPTION

[0042] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention are described clearly and completely below. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts are within the scope of protection of the present invention. In order to avoid obscuring the essence of the present invention, well-known methods and processes are not described in detail.

[0043] Example

[0044] The following combination Figures 1 to 3 , introduces an embodiment of the present invention: comprising the following specific steps:

[0045] Step 1: Modeling the network congestion dynamic graph structure {G t |G t =(V,A t )}.

[0046] The total sampling time n is divided into m time slices of equal length l. The devices in the network that can independently send and receive data are regarded as nodes and are uniquely identified. For example, device i corresponds to node v i , thus forming a node set V = {v i ,i=1,2,…,N v}, where N v is the total number of nodes.

[0047] If there is a directly connectable data transmission path between device i and device j, and the normalized data volume transmitted from i to j in time slice t is s ij,t .

[0048] Defining the network dynamic association matrix The subscript t indicates the value within time slice t.

[0049] According to the inherent links of the network, define the inherent adjacency matrix of the network

[0050]

[0051] In summary, within a specific time slice t, the connectivity status between nodes can be expressed by the network dynamic adjacency matrix A t Characterization, definition for:

[0052] A t =S t +A o , where element A ij,t =s ij,t +a ij .

[0053] Step 2: Screen candidate congestion propagation path nodes, candidate congestion propagation path set where N p is the total number of candidate congested propagation paths, and the adjacency matrix A of the participating nodes in the candidate congested paths is obtained. o- .

[0054] Generate candidate congestion propagation path set P and adjacency matrix A of participating nodes o- as follows:

[0055] According to the dynamic graph structure modeled in step 1, the directed connections between adjacent nodes in the time direction and adjacent time slices can represent the actual data transmission path in the network, such as Figure 1 .

[0056] The candidate congestion propagation path set is a set of nodes that are adjacent to each other. A candidate congestion path p can be represented as:

[0057] p=(v r1 ,v r2 ,…,v rd )

[0058] Set the path length d, Considering the calculation time, it is generally set to an integer less than 6 and greater than 3.

[0059] Calculate the current node v i The in-degree sequence:

[0060]

[0061] Take the maximum in-degree max{D i,t}Corresponding time slice t D The adjacent time slice t D-1 Upper v i All adjacent nodes of the current node are connected to generate the spatiotemporal neighboring point set of the current node

[0062] Similarly, calculate its in-degree sequence

[0063] Calculate and compare the current node v i Adjacent nodes in the spatiotemporal neighborhood Pearson correlation coefficient between in-degree sequences

[0064] Take the two most relevant adjacent nodes corresponding to the largest Pearson coefficient values ​​and generate two new paths. The two selected adjacent nodes are used as the previous nodes of the current node in the two new paths. Let the selected adjacent nodes be the new current node.

[0065] Repeat the above process of finding the two most relevant adjacent nodes on adjacent time slices based on the maximum in-degree until each path reaches the preset number of single path nodes d.

[0066] Record all new paths generated in the process to filter out all candidate paths

[0067] From the network inherent adjacency matrix A o Remove the corresponding values ​​of the nodes that do not participate in the construction of P and obtain the adjacency matrix A of the participating nodes o- .

[0068] Step 3: The node historical congestion state value corresponding to the candidate congestion propagation path {O rk,tk} is used as input and a vector representation H of the congestion propagation path is generated through a multi-layer LSTM module.

[0069] If a node v i There are N candidate congested paths in the set p paths, where each candidate congestion propagation path corresponds to a layer of LSTM network.

[0070] For a candidate path p r =(v r1 ,v r2 ,…,v rd ), the corresponding congestion propagation time slice is T r =(t r1 ,t r2 ,…,t rd ). Then the candidate path p of this node r In the corresponding LSTM network, the input of the kth LSTM unit is p r Medium v rk In time slice t rk The node congestion state value O rk,tk , and the state h output by the k-1th LSTM unit r,k-1 .

[0071] The state quantity h finally output by the rd LSTM unit r,rd As the path p r The eigenvalue of . Then N p The characteristic values ​​of the paths constitute the candidate path vector representation h of the node i .

[0072] Concatenate the candidate path vector representations of all participating network congestion nodes into a feature matrix N is the total number of nodes involved in network congestion.

[0073] Step 4: The graph convolution module outputs the node congestion prediction value, and the congestion propagation path vector H obtained in step 3 and the adjacency matrix A of the participating nodes in the candidate congestion path are represented. o- , as input through a graph convolution module consisting of a two-layer graph convolutional network, and then connected to a fully connected layer to adjust the dimension of the final output data, and the candidate path corresponding to the maximum value in the output result is selected as the predicted network congestion path.

[0074] The neural network model directly connected to the LSTM and graph convolutional network is based on the TensorFlow framework and is optimized using mean squared error and a common loss function:

[0075]

[0076] where Y and The actual value and the predicted value, the l2 regularization term avoids overfitting, λ is a learnable hyperparameter, L reg Represents the set of all parameters.

[0077] Based on experience, some hyperparameters were manually adjusted and set. For example, the learning rate was set to 0.00001, the number of training epochs was set to 1000, the L2 regularization coefficient λ was set to 0.05, and the output dropout rating of the LSTM network was set to 0.05. The optimal number of hidden units for the included graph convolutional network and LSTM network was obtained through grid search. For example, the number of hidden units in the LSTM network was set to 64, and the hidden units of the two-layer graph convolutional network were set to 32 and 16, respectively.

[0078] The present invention models the network topology and congestion status as graph structure data with both spatial and time-varying characteristics, and combines data feature mining and recurrent neural network prediction to predict the congestion propagation path that causes congestion at the target node.

[0079] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A congestion propagation path prediction method based on network information mining, characterized in that The steps include: Step 1: Model the data transmission network structure and its congestion state as a dynamic graph structure {G t |G t =(V,A t )}; The independent devices participating in network data transmission are the node set V in the figure = {v i ,i=1,2,…,N v }, where N v is the total number of nodes, and the connectivity state between nodes in time slice t is the network dynamic adjacency matrix A t ,The graphs of adjacent time slices can be connected according to the connectivity of data transmission paths in the network; Step 2: Screening candidate congestion propagation path node sets: candidate congestion propagation path node sets where N p is the total number of candidate congestion propagation paths, and the adjacency matrix A of the nodes participating in the network congestion in the candidate path set is obtained. o- ; Specifically, based on the dynamic graph structure modeled in step 1, the directed connections between adjacent nodes along the time direction between adjacent time slices can represent the actual data transmission path in the network; a candidate congested path p is represented by a set of nodes that are adjacent to each other: p=(v r1 ,v r2 ,…,v rd ) If v i is the terminal node v of the congestion propagation path rd , calculate v i The in-degree sequence: Take the maximum in-degree max{D i,t }Corresponding time slice t D The adjacent time slice t D-1 Upper v i All adjacent nodes of the current node are connected to generate the spatiotemporal neighboring point set of the current node Calculate its in-degree sequence Calculate and compare the current node v i Its adjacent nodes Pearson correlation coefficient between in-degree sequences Take the two adjacent nodes with the largest Pearson coefficient values ​​and generate two new paths. The two adjacent nodes are taken as the upper level nodes of the current node, and then the selected adjacent nodes are set as the new current node. Repeat the above process of finding the two most relevant adjacent nodes on adjacent time slices by the maximum in-degree until each path reaches the preset number of single path nodes Filter all candidate paths N p =2 d ; From the network inherent adjacency matrix A o Remove the corresponding values ​​of the nodes that do not participate in the construction of P and obtain the adjacency matrix A of the participating nodes o- ; Step 3: The historical congestion status value of the nodes participating in the network congestion propagation {O rk,tk } as input, generate the congestion propagation path vector representation H through the multi-layer LSTM module; Step 4: Represent the congestion propagation path vector H and the adjacency matrix A of the nodes involved in the network congestion o- , as input through a graph convolution module consisting of a two-layer graph convolutional network and a fully connected layer, and the path corresponding to the top d nodes in the output congestion status prediction value ranking is selected as the predicted network congestion path.

2. A congestion propagation path prediction method based on network information mining according to claim 1, characterized in that include: The dynamic graph structure reflecting the network structure and its congestion status in step 1; Its dynamic adjacency matrix is ​​defined as follows: The total sampling time n is divided into m time slices of equal length l. The devices in the network that can independently send and receive data are regarded as nodes and are uniquely identified. Device i corresponds to node v i If there is a directly connectable data transmission path between device i and device j, and the normalized data volume transmitted from i to j in time slice t is s ij,t , define the network dynamic association matrix The subscript t indicates the value in time slice t; Define the network intrinsic adjacency matrix In a specific time slice t, the network dynamic adjacency matrix A between nodes t for: A t =S t +A o 。 3. The method for predicting congestion propagation paths based on network information mining according to claim 1, characterized in that In step 3, the congestion propagation path vector representation H is generated by a multi-layer LSTM module as follows: If a node v i There are N candidate congestion propagation paths in the set p paths, where each candidate congestion propagation path corresponds to a layer of LSTM network; for a candidate congestion propagation path p r =(v r1 ,v r2 ,…,v rd ), the corresponding congestion propagation time slice is T r =(t r1 ,t r2 ,…,t rd ); then the candidate congestion propagation path p of the node r In the corresponding LSTM network, the input of the kth LSTM unit is p r Medium v rk In time slice t rk The node congestion state value O rk,tk , and the state h output by the k-1th LSTM unit r,k-1 ; The state quantity h finally output by the rd LSTM unit r,rd As the path p r The eigenvalue of p The characteristic values ​​of the paths constitute the candidate congestion propagation path vector representation h of the node i ; Concatenate the candidate congestion propagation path vectors of all participating network congestion nodes into a feature matrix N is the total number of nodes involved in network congestion.

4. The method for predicting congestion propagation paths based on network information mining according to claim 1, characterized in that The LSTM and graph convolutional network constitute a recurrent neural network model. The recurrent neural network model composed of LSTM and graph convolutional network is based on the Tensorflow framework and is optimized using mean square error and a universal loss function: where Y and The actual value and the predicted value, the l2 regularization term avoids overfitting, λ is a learnable hyperparameter, L reg Represents the set of all parameters.

Images