Multi-path routing optimization method and system based on graph attention network

Abstract

The invention relates to the field of routing algorithms of a mobile ad hoc network, discloses a multi-path routing optimization method and system for the mobile ad hoc network based on a graph attention network, and mainly solves the problem that in a traditional mobile ad hoc network, due to the fact that link characteristics and a routing scheme can generate nonlinear influence on network performance, the network performance is not influenced. And the problem of network resource waste caused by an improper routing scheme is solved. The content provided by the invention comprises a mobile ad hoc network multi-path network model, a graph attention network prediction model and a routing optimization system. The method comprises the following steps: firstly, generating path information through feature vectors of nodes and links obtained by a software defined network technology; secondly, constructing a multi-path network model of the mobile ad hoc network according to the hidden state relationship between the links and the paths; and finally, constructing a graph attention network prediction model in an application layer, thereby realizing the mobile ad hoc network multipath routing optimization method based on the graph attention network. In addition, a cross-layer routing optimization system is established according to the mobile ad hoc network multipath routing optimization method based on the graph attention network. According to the mobile ad hoc network multipath routing optimization method and system based on the graph attention network provided by the invention, node aggregation is optimized, the requirements of simultaneously learning network features and network complex structures in the mobile ad hoc network are met, the accuracy of predicting the network performance is improved, the routing strategy can be selected in an auxiliary manner, and the routing efficiency is improved. And the utilization rate of network resources is effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of routing algorithms for mobile ad hoc networks, and in particular to a method and system for optimizing multipath routing for mobile ad hoc networks based on a graph attention network. Background Art

[0002] With the development of the Internet era, the network communication performance problem of mobile ad hoc networks has become increasingly prominent. Mobile ad hoc networks are mobile networks in which wireless mobile nodes self-organize and network to complete node communication. The difference between mobile ad hoc networks and ordinary wireless networks is that they do not rely on wired infrastructure such as base stations to achieve mobile communication. Their characteristics of being centerless, self-organized, dynamic topology and multi-path make them often used in the fields of Internet of Things, Internet of Vehicles, drone clusters, wireless sensor networks, etc. In mobile ad hoc networks, changes in network topology due to changes in node positions, the addition and loss of nodes often occur. Therefore, the choice of routing protocols directly affects the communication performance of mobile ad hoc networks, which is also one of the key issues currently facing mobile ad hoc networks. Limited by the energy of wireless communication nodes in mobile ad hoc networks, source nodes generally cannot directly establish communication with destination nodes, and need to transmit data through multi-hop transmission forwarding through other nodes in the network. Therefore, routing protocols are very important in mobile ad hoc networks. According to the path selection of the routing protocol, it can be divided into single-path routing and multi-path routing. Among them, multi-path routing can select different routing protocols according to different network conditions to improve the stability and resource utilization of mobile ad hoc networks with unstable topology structures.

[0003] In order to select the most suitable routing solution for the current network environment, it is necessary to obtain all the node and link information on the network, which is difficult to do with conventional methods, but software-defined networks can do this. Software-defined networks allow us to control the network programmatically, meet the requirements of frequent link switching and routing strategy changes, and make it possible to maximize the use of network resources.

[0004] How to choose the best routing solution is the key issue to maximize the use of network resources and improve network performance, and the relationship between routing solutions and network performance is difficult to describe with simple modeling. In recent years, graph neural networks have performed well in recommendation systems, chemistry, medicine, and transportation because they can process data that is difficult for traditional neural networks to process, such as social network graphs, molecular graph structures, and transportation networks. Mobile ad hoc networks can be viewed as graph structures with mobile nodes as nodes and communication links as edges, so graph neural networks can be used to explore the relationship between routing solutions and network performance. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a method and system for optimizing multipath routing in a mobile ad hoc network based on a graph attention network in order to solve the defects involved in the background technology. The method first generates path information by obtaining feature vectors of nodes and links through software-defined network technology; secondly, constructs a multipath network model of a mobile ad hoc network according to the hidden state relationship between links and paths; finally, constructs a graph attention network prediction model at the application layer to implement a multipath routing optimization method for a mobile ad hoc network based on a graph attention network; in addition, a cross-layer routing optimization system is established according to the multipath routing optimization method for a mobile ad hoc network based on a graph attention network. The method and system for optimizing multipath routing in a mobile ad hoc network based on a graph attention network provided by the present invention optimize node aggregation, meet the requirements of learning network characteristics and complex network structures in a mobile ad hoc network at the same time, improve the accuracy of predicting network performance, can assist in selecting routing strategies, and effectively improve the utilization rate of network resources.

[0006] The present invention adopts the following technical solutions:

[0007] A multipath routing optimization method for mobile ad hoc networks based on graph attention network, characterized by comprising:

[0008] S1: Generate path information using feature vectors of nodes and links obtained through software-defined networking technology;

[0009] S2: Construct a multi-path network model for mobile ad hoc networks based on the hidden state relationship between links and paths;

[0010] S3: Build a graph attention network prediction model at the application layer and output a comprehensive score of network performance.

[0011] Specifically, in step S1, the node and link characteristics include bandwidth bw i , delay i , packet loss rate lo i 、Power pw i The link itself has no power pw i However, the stability of the link is limited by the nodes with lower power at both ends of the link. Therefore, the lower power of the nodes at both ends of the link is used as the link power pw i Path p i It is a path connecting the source node and the destination node, which can be expressed as:

[0012]

[0013] where p i is the i-th path of the routing solution, p ji is the path p i The jth link of i | is the path p iThe length of the path can be expressed as bw pi , pi ,lo pi , pw pi . p i The delay is p i The sum of all link delays is

[0014]

[0015] p i The bandwidth is p i The minimum value of all link bandwidths on

[0016]

[0017] p i The packet loss rate can be obtained by calculating the sum of the probability of no packet loss in the link, that is,

[0018]

[0019] p i The charge is p i The minimum value of the power of all nodes on the network is also the minimum value of the power of all links, that is,

[0020]

[0021] Specifically, in step S2, the hidden state of the link includes information on delay, packet loss rate, and link utilization, and the hidden state of the path includes information on response time, network throughput, and packet loss rate, and both include network structure information.

[0022] There are several paths in multipath routing. Obviously, the state of a path is determined by all the links that the path passes through, and the link is affected by the previous link in each path. Define the hidden state of the link as h l , define the hidden turntable of the path as h p , the relationship formula between the link and path hidden states is expressed as:

[0023]

[0024]

[0025] Among them, l i is the path p k The jnth link of Influence, and p i is affected by all the links it passes through. f() and g() are functions that describe the potential relationship between links and paths, respectively.

[0026] Furthermore, in step S3, the graph attention network prediction model regards links as nodes and paths as special nodes. The path and link affiliation is expressed as an ordered link sequence, and the same links in different paths have the same unique link number. The input of the model includes the link information of the network topology obtained by the software-defined network and the path information of the routing scheme. Since the paths may overlap, the links are uniquely numbered and the link information is stored in x l middle:

[0027]

[0028] Among them, de1, bw1, lo1, and pw1 are the delay, bandwidth, packet loss rate, and power of the first link respectively, and |L| is the number of all links.

[0029] Similarly, the path information is stored in x p middle:

[0030]

[0031] The path of the routing solution is composed of a sequence of link numbers. Since the path lengths are inconsistent, the path lengths need to be padded to form the matrix x r :

[0032]

[0033] The network routing is converted into a path-link graph, and the long short-term memory neural network (LSMT) of supervised learning is used to aggregate the link information into the path information to calculate the hidden state of the path. Then, according to the mutual influence relationship between the links, the attention mechanism is used to calculate the weights to aggregate the link information and calculate the hidden state of the link. The above process is repeated many times, the parameters are continuously optimized, and finally the hidden state of the link and path is input into the multi-layer perceptron (MLP) to obtain the output.

[0034] The specific process of graph attention network prediction is as follows:

[0035] Step 1: Parameter initialization;

[0036] Step 2: Input information matrix x l ,x p ,x r ;

[0037] Step 3: Hidden states h of links and paths l and h p Initialize with input x l Matrix and x p The matrix fills the corresponding part and adds 0 to lengthen it;

[0038] Step 4: Iterate the model;

[0039] Step 5: Path aggregation links are aggregated using LSMT, and the order relationship of the links is recorded at this stage to provide conditions for the calculation of the attention mechanism weights;

[0040] Step 6: The attention mechanism of the link calculates the weight by recording the sequence relationship of the links in each path;

[0041] Step 7: Calculate the hidden state of the link;

[0042] Step 8: Read out the data, and use the multilayer perceptron with activation function according to the hidden state h l and h p The characteristic vector of the graph is obtained, and the network throughput and stability as well as other network performance indicators are predicted by the difference in data labels.

[0043] A multipath routing optimization system for mobile ad hoc networks based on graph attention networks adopts a multipath routing optimization method for mobile ad hoc networks based on graph attention networks. The system includes two parts: a mobile ad hoc network and a software-defined network controller. The software-defined network separates the control plane from the data plane, realizes network programmability, provides a global view of the entire network, and uses the openflow routing protocol to control connections.

[0044] Furthermore, the software-defined network controller includes a network information detection module, a routing exploration module, a graph attention network model, and a routing decision module.

[0045] The network information detection module is the key module that connects the mobile ad hoc network and the application. When the mobile ad hoc network node initiates a transmission request, the software-defined network collects network information and updates the link information matrix x l , and update the transmission request table at the same time. The software-defined network needs to actively update the link information table and transmission request table every once in a while.

[0046] The routing exploration module provides the path link matrix x for the model output r ,When the link information matrix and the transmission request table are updated, the ,routing exploration module generates a routing scheme consisting of a set of paths from the ,source node to the target node based on the classic routing algorithm, ,Dijkstra algorithm, A-Star algorithm and random algorithm. Then the ,alternative routing set is input into the graph attention ,network model for network performance prediction and as a reference for ,routing decision.

[0047] The graph attention network model is the core part of the entire routing optimization system. In the offline training phase, labeled samples with different numbers of nodes, different routing algorithms, and different scenarios are collected to train the graph attention network model. In the online use phase, the candidate routing set of the routing exploration module is sequentially input into the graph attention network model for network performance prediction. Based on the prediction results, the network performance score is obtained and the score result is sent to the routing decision module. The model predicts the throughput and stability of the network respectively, and calculates the network score accordingly. The formula is:

[0048]

[0049] Where NP is the comprehensive network score, P is the set of routing paths, |P| is the number of paths, and p i is the i-th path, R pi is the path p i The stability of y(p i ) is the path p i throughput, k is the expected performance loss coefficient, which is set to 0.5.

[0050] The routing decision module is the only channel for the software-defined network to control the real network. After receiving the scoring results sent by the graph attention network model, the routing decision module sends the configuration to the nodes in the network according to the optimal routing solution, and the nodes send data packets according to the configuration information.

[0051] Compared with the prior art, the present invention adopts the above technical solution and has the following beneficial effects:

[0052] 1. The present invention establishes a multipath routing optimization method and system for mobile ad hoc networks based on graph attention networks, which mainly solves the problem that in traditional mobile ad hoc networks, inappropriate routing schemes lead to waste of network resources due to the nonlinear impact of link characteristics and routing schemes on network performance.

[0053] 2. The present invention uses software-defined networking technology to separate the control plane and data plane in the communication system. In the control layer, the controller masters the topology information of the global network and is responsible for the configuration management and protocol deployment in the network, while in the data layer, the network nodes are only responsible for fast forwarding of data packets. Software-defined networking allows us to control the network in a programmatic way, meet the requirements of frequent link switching and routing strategy changes, and maximize the use of network resources.

[0054] 3. The present invention takes the link information of the network topology and the path information of the routing scheme as input, and the predicted network throughput and stability and other network performance indicators as output, and constructs a graph attention network model at the application layer, which can further optimize the node aggregation process. The graph attention network meets the requirements of the model to learn both network features and complex network structures at the same time, improves the accuracy of predicting network performance, and can assist in selecting routing strategies.

[0055] 4. The present invention establishes a cross-layer routing optimization system based on a multipath routing optimization method for mobile ad hoc networks based on a graph attention network. The system includes a mobile ad hoc network and a software defined network controller. The SDN controller includes a network information detection module, a routing exploration module, a graph attention network prediction model module, and a routing decision module. The multipath routing optimization system for mobile ad hoc networks based on a graph attention network can significantly improve throughput and network stability on multi-node high-load mobile ad hoc networks. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 It is a schematic diagram of network graph model transformation of the present invention;

[0057] Figure 2 It is a schematic diagram of the graph attention network prediction model of the present invention;

[0058] Figure 3 It is a schematic diagram of a multipath routing optimization system for a mobile ad hoc network based on a graph attention network according to the present invention. DETAILED DESCRIPTION

[0059] The technical solution of the present invention is further described in detail below in conjunction with the accompanying drawings:

[0060] 1) If Figure 1 As shown in the figure, graph neural networks process graph data composed of nodes and links. Generally, nodes carry the main features, and links have an impact in the form of adjacency matrix and node aggregation. In the wireless ad hoc network model, links carry the main features, and the relationship between links and paths plays a major role in learning graph structures. Figure 1 As shown, the actual routing graph and path-link relationship graph are as follows: Figure 1 In (a), 1, 5, 2, 3 and 4, 5, 6 and 7 are paths respectively. Figure 1 In (b), links are regarded as nodes and paths are regarded as special nodes. For example, the path composed of links 1, 5, 2, and 3 is regarded as node 9. The path and link affiliation are shown in the figure as an ordered link sequence. The same links in different paths have the same unique link sequence number.

[0061] The input of the model includes the link information of the network topology obtained by the software-defined network and the path information of the routing solution. Since the paths may overlap, the links are uniquely numbered and the link information is stored in x l middle:

[0062]

[0063] Among them, de1, bw1, lo1, and pw1 are the delay, bandwidth, packet loss rate, and power of the first link respectively, and |L| is the number of all links.

[0064] Similarly, the path information is stored in x p middle:

[0065]

[0066] The path of the routing solution is composed of a sequence of link numbers. Since the path lengths are inconsistent, the path lengths need to be padded to form the matrix x r :

[0067]

[0068] 2) If Figure 2 As shown in the figure, the graph attention network prediction model transforms the network routing into a path-link graph, uses the supervised learning long short-term memory neural network (LSMT) to aggregate the link information into the path information to calculate the hidden state of the path, and then uses the attention mechanism to calculate the weights to aggregate the link information and calculate the hidden state of the link based on the mutual influence relationship between the links. The above process is repeated many times, the parameters are continuously optimized, and finally the hidden state of the link and path is input into the multi-layer perceptron (MLP) to obtain the output.

[0069] The specific process of graph attention network prediction is as follows:

[0070] Step 1: Parameter initialization;

[0071] Step 2: Input information matrix x l ,x p ,x r ;

[0072] Step 3: Hidden states h of links and paths l and h p Initialize with input x l Matrix and x p The matrix fills the corresponding part and adds 0 to lengthen it;

[0073] Step 4: Iterate the model;

[0074] Step 5: Path aggregation links are aggregated using LSMT, and the order relationship of the links is recorded at this stage to provide conditions for the calculation of the attention mechanism weights;

[0075] Step 6: The attention mechanism of the link calculates the weight by recording the link sequence relationship in each path;

[0076] Step 7: Calculate the hidden state of the link;

[0077] Step 8: Read out the data, and use the multilayer perceptron with activation function according to the hidden state h l and h p The characteristic vector of the graph is obtained, and the network throughput, stability and other network performance indicators are predicted by the difference in data labels.

[0078] 3) If Figure 3 As shown in the figure, the multipath routing optimization system for mobile ad hoc networks based on graph attention networks consists of two parts: a mobile ad hoc network and a software-defined network controller. The software-defined network separates the control plane from the data plane to achieve network programmability. The software-defined network provides a global view of the entire network and uses the OpenFlow routing protocol to control connections. The software-defined network controller includes a network information detection module, a routing exploration module, a graph attention network prediction model module, and a routing decision module.

[0079] The network information detection module is the key module that connects the mobile ad hoc network and the application. When the mobile ad hoc network node initiates a transmission request, the software-defined network collects network information and updates the link information matrix x l , and update the transmission request table at the same time. The software-defined network needs to actively update the link information table and transmission request table every once in a while.

[0080] The routing exploration module provides the path link matrix x for the model output r ,When the link information matrix and the transmission request table are updated, the ,routing exploration module generates a routing scheme consisting of a set of paths from the ,source node to the target node based on the classic routing algorithm, ,Dijkstra algorithm, A-Star algorithm and random algorithm. Then the ,alternative routing set is input into the graph attention ,network model for network performance prediction and as a reference for ,routing decision.

[0081] The graph attention network model is the core part of the entire routing optimization system. In the offline training phase, labeled samples with different numbers of nodes, different routing algorithms, and different scenarios are collected to train the graph attention network model. In the online use phase, the candidate routing set of the routing exploration module is sequentially input into the graph attention network model for network performance prediction. Based on the prediction results, the network performance score is obtained and the score result is sent to the routing decision module. The model predicts the throughput and stability of the network respectively, and calculates the network score accordingly. The formula is:

[0082]

[0083] Where NP is the comprehensive network score, P is the set of routing paths, |P| is the number of paths, and p i is the i-th path, R pi is the path p i The stability of y(p i ) is the path p i throughput, k is the expected performance loss coefficient, which is set to 0.5.

[0084] The routing decision module is the only channel for the software-defined network to control the real network. After receiving the scoring results sent by the graph attention network model, the routing decision module sends the configuration to the nodes in the network according to the optimal routing solution, and the nodes send data packets according to the configuration information.

[0085] The present invention is applicable to route optimization of mobile ad hoc networks. By utilizing a multipath route optimization method and system for mobile ad hoc networks based on a graph attention network, throughput and network stability can be significantly improved on multi-node high-load mobile ad hoc networks, the accuracy of predicting network performance can be improved, routing strategies can be assisted in selecting, and the problem of network resource waste caused by inappropriate routing solutions can be solved.

[0086] It will be understood by those skilled in the art that, unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as those generally understood by those skilled in the art in the art to which the present invention belongs. It should also be understood that terms such as those defined in common dictionaries should be understood to have meanings consistent with the meanings in the context of the prior art, and will not be interpreted with idealized or overly formal meanings unless defined as herein.

[0087] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A multipath routing optimization method for mobile ad hoc networks based on graph attention network, characterized in that: include: S1: Generate path information using feature vectors of nodes and links obtained through software-defined networking technology; S2: Construct a multi-path network model for mobile ad hoc networks based on the hidden state relationship between links and paths; S3: Build a graph attention network prediction model at the application layer and output a comprehensive score of network performance.

2. According to claim 1, the multipath routing optimization method for mobile ad hoc networks based on graph attention network is characterized in that: The node and link characteristics in step S1 include bandwidth bw i , delay i , packet loss rate lo i 、Power pw i ; The link itself has no power pw i However, the stability of the link is limited by the nodes with lower power at both ends of the link. Therefore, the lower power of the nodes at both ends of the link is used as the link power pw i Path p i It is a path connecting the source node and the destination node, which can be expressed as: where p i is the i-th path of the routing solution, p ji is the path p i The jth link of i | is the path p i The length of the path can be expressed as bw pi , pi ,lo pi , pw pi ;p i The delay is p i The sum of all link delays is p i The bandwidth is p i The minimum value of all link bandwidths, that is, p i The packet loss rate can be obtained by calculating the sum of the probability of no packet loss in the link, that is, p i The charge is p i The minimum value of the power of all nodes on the network is also the minimum value of the power of all links, that is, 3. The multipath routing optimization method for mobile ad hoc networks based on graph attention network according to claim 1 is characterized in that: In step S2, the hidden state of the link includes information about delay, packet loss rate, and link utilization, and the hidden state of the path includes information about response time, network throughput, and packet loss rate, and both include network structure information; There are several paths in multipath routing. Obviously, the state of a path is determined by all the links that the path passes through, and the link is affected by the previous link in each path. The hidden state of the link is defined as h l , define the hidden turntable of the path as h p , the relationship formula between the link and path hidden states is expressed as: Among them, l i is the path p k The jnth link of Influence, and p i is affected by all the links it passes through; f() and g() are functions that describe the potential relationship between links and paths, respectively.

4. The multipath routing optimization method for mobile ad hoc networks based on graph attention network according to claims 1-3 is characterized in that: The graph attention network prediction model described in step S3 regards links as nodes and paths as special nodes. The path and link affiliation is expressed as an ordered link sequence, and the same links in different paths have the same unique link number. The input of the model includes the link information of the network topology obtained by the software-defined network and the path information of the routing scheme. Since the paths may overlap, the links are uniquely numbered and the link information is stored in x l middle: Where de1, bw1, lo1, and pw1 are the delay, bandwidth, packet loss rate, and power of the first link, respectively, and |L| is the number of all links; Similarly, the path information is stored in x p middle: The path of the routing solution is composed of a sequence of link numbers. Since the path lengths are inconsistent, the path lengths need to be padded to form the matrix x r : The network routing is converted into a path-link graph, and the long short-term memory neural network (LSMT) of supervised learning is used to aggregate the link information into the path information to calculate the hidden state of the path. Then, according to the mutual influence relationship between the links, the attention mechanism is used to calculate the weights to aggregate the link information and calculate the hidden state of the link. The above process is repeated multiple times, the parameters are continuously optimized, and finally the hidden state of the link and path is input into the multi-layer perceptron (MLP) to obtain the output.

5. A multipath routing optimization system for mobile ad hoc networks based on graph attention network, characterized in that: include: Mobile ad hoc networks and software-defined network controllers; Software-defined networking separates the control plane from the data plane, implements network programmability, provides a global view of the entire network, and uses the OpenFlow routing protocol to control connections; The software-defined network controller includes a network information detection module, a routing exploration module, a graph attention network model, and a routing decision module; The network information detection module is the key module that connects the mobile ad hoc network and the application. When the mobile ad hoc network node initiates a transmission request, the software-defined network collects network information and updates the link information matrix, and also updates the transmission request table; Software-defined networks need to actively perform link information tables and transmission request tables at regular intervals; The routing exploration module provides a path link matrix for the model output. When the link information matrix and the transmission request table are updated, the routing exploration module generates a routing solution consisting of a set of paths from the source node to the target node based on the classic routing algorithm, Dijkstra algorithm, A-Star algorithm and random algorithm; Then the candidate routing set is input into the graph attention network model to predict network performance and serve as a reference for routing decisions; The graph attention network model is the core part of the entire routing optimization system; In the offline training phase, labeled samples with different numbers of nodes, different routing algorithms, and different scenarios are collected to train the graph attention network model; In the online use phase, the candidate routing set of the routing exploration module is sequentially input into the graph attention network model for network performance prediction. Based on the prediction results, the network performance score is obtained and the score result is sent to the routing decision module. The model predicts the throughput and stability of the network respectively and calculates the network score accordingly. The formula is: Where NP is the comprehensive network score, P is the set of routing paths, |P| is the number of paths, and p i is the i-th path, R pi is the path p i The stability of y(p i ) is the path p i The throughput of , k is the expected performance loss coefficient, which is set to 0.5; The routing decision module is the only channel for the software-defined network to control the real network. After receiving the scoring results sent by the graph attention network model, the routing decision module sends the configuration to the nodes in the network according to the optimal routing solution, and the nodes send data packets according to the configuration information.

Images