Dynamic network routing optimization method and system based on deep learning prediction
By combining deep learning methods with pseudo-labels and contrastive learning, a network performance prediction model is optimized, which solves the problems of network complexity and high cost in existing technologies and achieves reliable real-time routing optimization and performance prediction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- EAST CHINA NORMAL UNIV
- Filing Date
- 2023-12-14
- Publication Date
- 2026-07-10
AI Technical Summary
Existing network performance prediction methods fail to effectively consider the complexity of real-world networks and the high cost of model training, resulting in the inability to provide reliable real-time routing optimization results to address network degradation and performance bottlenecks.
We employ a deep learning-based approach that combines weakly supervised pseudo-labels and contrastive learning of positive sample pairs. Through pseudo-label generation and model pre-training, we fine-tune the model using a small number of real-world labels to optimize the transfer learning capability of the network performance prediction model and achieve reliable route optimization.
It improves the accuracy and generalization ability of network performance prediction, reduces data collection costs, and enables real-time network performance prediction and routing optimization.
Smart Images
Figure CN117714307B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of network routing technology, and in particular to a dynamic network routing optimization method based on a deep learning prediction model. Background Technology
[0002] With the continuous development of communication technology, large-scale heterogeneous networks are gradually becoming a key part of future communication. Accurate network modeling allows us to understand the network's operating mechanisms and predict its behavior, such as key performance indicators (KPIs) like traffic distribution and transmission latency. This enables targeted network optimization, improving operational efficiency and reducing the probability of failures. However, in practical applications, the complex and ever-changing network environment, especially in large-scale satellite networks (such as Starlink), presents significant challenges to traditional network modeling and optimization methods due to the dynamic nature of satellite interactions and the complex ground-space link conditions.
[0003] Currently, many attempts have been made to develop accurate transmission models for networks. The paper [1] "V. Arun and H. Balakrishnan, "Copa: Practical delay-based congestion control for the internet," in 15th USENIX Symposium on Networked Systems Design and Implementation (NSDI 18), 2018, pp. 329–342" proposes to use discrete-state Markov models and queuing theory to model the optimal transmission rate as a function of queuing delay. However, this method usually assumes that the arrival of data packets follows a Poisson distribution, which is inaccurate in wide area networks. The paper [2] "K.Rusek, J.Su′arez-Varela, P.Almasan, P.Barlet-Ros, and A.CabellosAparicio, "Routenet: Leveraging graph neural networks for network modeling and optimization in SDN", IEEE Journal on Selected Areas in Communications, vol.38, no.10, pp.2260–2270, 2020. uses a heterogeneous graph neural network model to model network topology, effectively predict network performance, and perform route upgrades and optimizations. However, it suffers from expensive data acquisition costs and weak generalization.
[0004] In summary, current common network performance prediction methods do not take into account the complexity of real-world networks and the high cost of model training, making it impossible to provide reliable real-time prediction results for effective route optimization to address network degradation and performance bottlenecks. Summary of the Invention
[0005] To address the shortcomings of existing technologies, the present invention aims to provide a dynamic network routing optimization method and system based on deep learning prediction. By introducing weakly supervised pseudo-labels and contrastive learning positive sample pairs, the transfer learning capability of the model in the pre-training stage is optimized. Furthermore, a small number of real-world scene labels are used to fine-tune the model, thereby obtaining reliable network performance prediction results for routing optimization.
[0006] The dynamic network routing optimization method and system based on deep learning prediction provided by the present invention includes the following steps:
[0007] Step S1: For the end-to-end network traffic input in the network topology, the path traversed by the network traffic, the routing buffer queue and multi-hop link information, and the corresponding statistical indicators are combined to finally construct a graph network with heterogeneous nodes.
[0008] Step S2: Based on heterogeneous graph information, use the queuing theory (QT) algorithm to generate a large number of pseudo-labels for network traffic features related to latency and packet loss metrics (network performance metrics);
[0009] Step S3: Construct a network performance prediction model and pre-train the model based on heterogeneous graph networks and network traffic feature data with pseudo-labels;
[0010] Step S4: Based on the pre-trained network performance prediction model, fine-tuning training of the model is carried out using heterogeneous graph networks and a small amount of network traffic feature data with real labels.
[0011] Step S5: Using the network prediction model trained in the above steps, compare different routing methods to obtain the performance indicators of each traffic flow in the current network, and select the better-performing solution as the optimized routing solution.
[0012] The network performance prediction model is based on the GNN (Graph Transformer) model and MLP module in the pre-training stage, only the pre-trained GNN model is used in the fine-tuning stage, and the fine-tuned GNN model is used in the route optimization stage.
[0013] Compared with the prior art, the present invention has the following significant advantages:
[0014] (1) This invention designs a modeling method based on attention mechanism to characterize the long-term temporal dependency relationship between multiple data streams in the network topology, thereby improving its representation learning ability.
[0015] (2) This invention proposes a weakly supervised model training paradigm, which generates a large amount of high-quality pseudo-label data for model pre-training by means of queuing theory and other methods.
[0016] (3) The present invention constructs contrastive features in model pre-training as coarse-grained supervision signals, which enhances the model’s transfer and generalization ability across scenarios and helps with small-sample learning for specific tasks. Attached Figure Description
[0017] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:
[0018] Figure 1 This is a schematic diagram of the dynamic network routing optimization method based on deep learning prediction proposed in this invention.
[0019] Figure 2 This is a schematic diagram of the network performance prediction model based on weak supervision pre-training and the downstream routing optimization task in an embodiment of the present invention.
[0020] Figure 3 This is the architecture of a network performance prediction model based on weakly supervised pre-training in an embodiment of the present invention.
[0021] Figure 4 This is a flowchart of the downstream routing optimization task in an embodiment of the present invention.
[0022] Figure 5 This is a block diagram of the dynamic network routing optimization system designed in this invention. Detailed Implementation
[0023] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the scope of protection of the present invention.
[0024] This invention provides a dynamic network routing optimization method and system based on deep learning prediction. It achieves pseudo-label generation of key network traffic performance indicators based on queuing theory, and constructs a network performance prediction model based on weakly supervised learning and Siamese network comparative learning using these pseudo-labels, completing model pre-training. Furthermore, it uses a small amount of real network traffic data and its performance indicators to fine-tune the model, enabling a network performance prediction model with strong generalization performance at a relatively low data acquisition cost. Based on this, a dynamic routing strategy is designed according to the reliable performance prediction results under different routing tables to achieve real-time routing optimization.
[0025] Combination Figure 1 and Figure 2 The present invention provides a dynamic network routing optimization method and system based on deep learning prediction, comprising the following steps:
[0026] Step S1: For the end-to-end network traffic input in the network topology, the path traversed by the network traffic, the routing buffer queue and multi-hop link information, and the corresponding statistical indicators are combined to finally construct a graph network with heterogeneous nodes.
[0027] Step S2: Based on heterogeneous graph information, use the queuing theory (QT) algorithm to generate a large number of pseudo-labels for network traffic features related to latency and packet loss metrics (network performance metrics);
[0028] Step S3: Construct a network performance prediction model and pre-train the model based on heterogeneous graph networks and network traffic feature data with pseudo-labels;
[0029] Step S4: Based on the pre-trained network performance prediction model, fine-tuning training of the model is carried out using heterogeneous graph networks and a small amount of network traffic feature data with real labels.
[0030] Step S5: Using the network prediction model trained in the above steps, compare different routing methods to obtain the performance indicators of each traffic flow in the current network, and select the better-performing solution as the optimized routing solution.
[0031] The network performance prediction model is based on the GNN (Graph Transformer) model and MLP module in the pre-training stage, only the pre-trained GNN model is used in the fine-tuning stage, and the fine-tuned GNN model is used in the route optimization stage.
[0032] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.
[0033] Example 1
[0034] The specific implementation method of the dynamic network routing optimization method based on deep learning prediction provided by the present invention is as follows:
[0035] Step S1.1: For the input network traffic, separate the terminal nodes (denoted as n) and end-to-end data streams (denoted as n) in the network traffic. ), the buffer queue of the routing node, and the one-hop link (denoted as ), As nodes on a heterogeneous graph, edges are formed between heterogeneous nodes based on the links traversed by the data flow and their corresponding relationships. Ultimately, the paths, links, and queue information traversed by each data flow in the network topology are organized into a heterogeneous graph network.
[0036] Step S1.2: Statistically record the traffic information of data packets in the network topology in multi-hop routing and global end-to-end according to a certain time interval, so as to represent the statistical information such as data packet throughput and bit rate at the current moment, thereby forming the traffic characteristics of the heterogeneous graph network;
[0037] The specific traffic characteristics are as follows:
[0038] Step S1.2.1: Record the routing and forwarding matrix (denoted as R) of the current network topology;
[0039] Given a network data flow with a fixed routing table, record the flow matrix R, where the element (i, j) indicates that node i and node j are directly reachable;
[0040] Step S1.2.2: Record the traffic matrix of the current network topology (denoted as P);
[0041] Considering the average bandwidth between each pair of nodes at the current moment, the total number of data packets, the distribution type of data packet arrival time and its distribution parameters, and the distribution type of data packet size and its distribution parameters, a dictionary structure is formed and stored in...
[0042] P i，j middle;
[0043] Step S1.2.3: Record the performance matrix of the current network topology (denoted as T);
[0044] The packet loss and delay quantile information between each pair of nodes at the current moment are used as T. i，j The element (the actual packet loss or delay indicator is marked as y);
[0045] Step S1.2.4: Record the static characteristics of each heterogeneous node and the edge types between heterogeneous nodes;
[0046] Record the characteristics of each node in the heterogeneous graph: terminal nodes include the size of the buffer queue (representing the number of data packets stored, which can be denoted as b). i Features include: ) and ); Link nodes include the bandwidth of the two nodes, scheduling policies (for QoS queues) and other features; Queue nodes include the size of the buffer queue, scheduling policies and other features; Path nodes include scheduling policy type, data flow service type (ToS), data packet arrival time distribution parameters, data packet size distribution parameters, average bandwidth, average number of data packets, latency and packet loss and other features; Record heterogeneous node features include the corresponding buffer queue size of the routing node, path length of the data flow and other static features.
[0047] Record the edge types in the heterogeneous graph: the edge type between a queue node and a path or link node is marked as -1; the edge type between a path node and a link node is marked as 0; the edge type between a path node and a terminal node is marked as 1; and the edge type between a link node and a terminal node is marked as 2.
[0048] Step S2.1: Obtain the network traffic routing relationship of each node at the current time step;
[0049] In the heterogeneous graph, select edges of type 0 that are composed of path nodes and link nodes. For each starting node, select the first set of edges, the second set of edges, and so on. Use these set of edges as different network traffic at each time step.
[0050] Step S2.2: Based on the traffic characteristics of the links and path nodes in Step 1.2, initialize the average traffic A of each path, the bandwidth capacity c of each link, the link traffic λ, and the link utilization ρ;
[0051] Step S2.3: Calculate the utilization rate and blocking probability of each link;
[0052] The link utilization rate between terminal node i and terminal node j at the current time step can be calculated by dividing the cumulative traffic of the link by the link bandwidth capacity, expressed as: Further calculation of the link blocking probability at the current time step, i.e., the probability of queuing delay occurring based on the buffer queue size, is expressed as:
[0053] Step S2.4: Update the link traffic λ according to the blocking probability and obtain the predicted packet loss for the current iteration;
[0054] The link traffic is initialized to empty. The first edge set is assigned the initial traffic A. Subsequent edge sets accumulate the traffic of the current path on the next hop link based on the blocking probability of each link, and the traffic is aggregated according to the corresponding edge index. The final link traffic λ is obtained by iterating through multiple hop links. Further representation is as follows;
[0055]
[0056]
[0057]
[0058]
[0059] Step S2.5: Continuously update the blocking probability until there is no absolute difference in the traffic of a link before and after two updates that is greater than the threshold of 0.001;
[0060] Step S2.6: Exit the loop and calculate the time delay prediction for each path node;
[0061] First, calculate the probability (denoted as π0) that the link queue is empty and its average utilization rate (denoted as π0). The delay of a given path can be obtained by aggregating the delays of all links along that path, which represents the cumulative queuing delay along that path.
[0062]
[0063]
[0064]
[0065] Step S3.1: Construct a network prediction model based on GNN message passing and attention, such as... Figure 3 As shown, the path sequence, combining queue features and link features, is first used to obtain the temporal representation of the path. Then, the path representation is used as the queue sequence to update the queue representation. Finally, the queue is used as the link sequence to update the link representation. Multiple iterations yield the feature representations of the path nodes.
[0066] Step S3.2: Construct the loss function for the pre-trained model, including fitting the QT pseudo-labels from the performance metric (denoted as ). The regression / classification loss, flow-level contrastive loss, and network-level contrastive loss are compared with the model's predicted values.
[0067] Step S3.3: Pre-train a network performance index model based on QT pseudo-label loss and same-sign comparison loss until convergence;
[0068] The specific model feature transfer process is as follows:
[0069] Step S3.1.1: Initialize feature representation;
[0070] The feature input of a given heterogeneous node is transformed into three feature representations of heterogeneous nodes. The path node features are initialized as average bandwidth and average number of packets (denoted as X). p The characteristics of a link node are the number of links, the link capacity, and the scheduling policy type (denoted as X). l The initialization of queue node characteristics includes queue size, scheduling policy, and weight (denoted as X). q );
[0071] Step S3.1.2: Update path features;
[0072] Choosing edge types of -1 and 0 in the heterogeneous graph represents the features of link and queue nodes that are related to path nodes. Learning these path features using a recurrent neural network (RNN) can be represented as RNN(concat([X...))l X q ], dim =
[0073] -1), X p ), where concat([X l X q ], dim = -1) indicates that the links and queues traversed by the path are concatenated according to the feature dimension and used as the input sequence; X p This represents the initial hidden state in the RNN. The new path state representation is obtained through the RNN update and used for subsequent reading function performance metric prediction.
[0074] Step S3.1.3: Update queue characteristics;
[0075] Obtain the path state representation X for the current iteration. p Then, it is repartitioned based on edges of type 0 to obtain the input path sequence about the queue, and X is... q The RNN is updated using the initial hidden state;
[0076] Step S3.1.4: Update link characteristics;
[0077] Similarly, it is repartitioned based on edges of type -1 to obtain the input queue sequence for the links, and X is... l The RNN is updated using the initial hidden state;
[0078] Step S3.1.5: Predict performance metrics;
[0079] After several iterations, the state representation X of each path is determined. p A self-attention module is used for global representation, enabling contextual interaction between paths. Then, an MLP module with ReLU activation is used to map path features to performance metric prediction results (denoted as...). );
[0080] The specific loss function is constructed as follows:
[0081] Step S3.2.1: Construct regression or classification losses for latency or packet loss metrics, using the MAPE loss function and the BCE loss function respectively.
[0082]
[0083]
[0084] Among them This refers to the pseudo-labels of the performance metrics generated based on QT in step S2. This represents the model's performance metric prediction results for the current network data;
[0085] Step S3.2.2: Construct the flow-level contrast loss;
[0086] The network topology and heterogeneous node features from two types of network datasets (which can be two types of network datasets from the same scenario or two types of network datasets from different scenarios) are input. Through step S3.1, performance prediction results for different data streams are obtained, denoted as... and Constructing the flow-level contrast loss is
[0087]
[0088]
[0089]
[0090] The method uses an MLP with ReLU and sigmoid activation functions to convert the difference between the predictions of the two networks into floating-point numbers between 0 and 1.
[0091] Step S3.2.3: Construct network-level contrastive loss;
[0092] Similar to step S3.2.3, but averaging the difference results and constructing a network-level contrastive loss by fitting regression values.
[0093]
[0094]
[0095]
[0096] Step S3.2.4: Construct the overall loss of the pre-trained model;
[0097] The model pre-training loss for latency performance metrics is:
[0098] loss delay =λ1loss mape +λ2loss flow +λ3loss net
[0099] The model pre-training loss for packet loss performance metrics is:
[0100] loss drop =λ1loss bce +λ2loss flow +λ3loss net
[0101] Step S4.1: Similar to the network prediction model based on GNN message passing and attention in step S3.1, the model provides prediction results for the given network topology and various traffic characteristics;
[0102] Step S4.2: Construct the loss function for model fine-tuning, which is the regression / classification loss that fits the true label (denoted as y) from the performance index to the model prediction value. The loss functions for latency and packet loss index can be expressed as follows;
[0103]
[0104]
[0105] Step S4.3: Fine-tune the network performance index model based on the real label loss performance index until convergence;
[0106] Example 2
[0107] The routing optimization method based on a network performance prediction model provided by this invention is implemented as follows:
[0108] Step S5.1: As Figure 4 As shown on the right side of the route optimization algorithm flowchart, firstly, for each origin node i and destination node j, a set of no more than k candidate paths is selected to update the routing table matrix R, i.e., Ri. i，j ={path1, ...,path k The candidate paths are selected based on the network topology and the path length algorithm between two points i and j, with the path length less than the shortest length plus a threshold (thresh), and the candidate path R is guaranteed to be... i，j The total number is less than or equal to the threshold k.
[0109] Step S5.2: As Figure 4 The left side of the routing optimization algorithm flowchart shows a table of SLA thresholds for different data flow types, a set of traffic that meets the SLA (denoted as Sat), and a set of traffic that does not meet the SLA (denoted as No_SAT).
[0110] Step S5.3: Further calculate the real-time network data flow latency under the given path in the current routing table based on the network performance prediction model, and combine the prediction result with the SLA threshold table. If the current traffic that meets the SLA increases, reset the iteration count, filter the current data flow and update Sat and No_Sat.
[0111] Step S5.4: Randomly select 1 / 5 of the traffic from both the Sat and No_Sat traffic sets, and randomly select one path from the candidate path set to form a new routing table for one iteration. Repeat step S5.3. A total of Iter is calculated, where Iter represents the number of iterations per round.
[0112] Step S5.5: Specify the number of rounds, repeat step S5.4, and if the number of rounds exceeds the specified number, return the final routing table to guide the transmission of data streams;
[0113] Example 3:
[0114] This invention also provides a dynamic network routing optimization system based on deep learning prediction, referring to... Figure 5 As shown, the system includes:
[0115] Network feature acquisition module M1: Based on multiple routing and forwarding devices under a specified topology, network traffic data at a certain time interval is collected and statistically analyzed to form a network traffic feature sequence according to the feature requirements of heterogeneous nodes in step S1.
[0116] Specifically, for the input network traffic, the terminal nodes (denoted as n) and the end-to-end data streams (denoted as n) in the network traffic are... ), the buffer queue of the routing node, and the one-hop link (denoted as ), As nodes on a heterogeneous graph, edges are formed between heterogeneous nodes based on the links traversed by the data flow and their corresponding relationships. Ultimately, the paths, links, and queue information traversed by each data flow in the network topology are organized into a heterogeneous graph network.
[0117] It also counts and records the traffic information of data packets in the network topology in multi-hop routing and global end-to-end according to a certain time interval, so as to represent the statistical information such as data packet throughput and bit rate at the current moment, thereby forming the traffic characteristics of heterogeneous graph networks;
[0118] Network performance prediction model pre-training module M2: Based on steps S2 and S3, pre-train the network performance prediction model using the QT algorithm and contrastive learning strategy;
[0119] Specifically, edges of type 0 formed by path nodes and link nodes in the heterogeneous graph are selected. For each starting node, a first set of edges, a second set of edges, and so on, are chosen, and these set sets are used as different network traffic at each time step. The QT algorithm is used to calculate the average traffic A of each path, the bandwidth capacity c of each link, the link traffic λ, and the link utilization ρ. Finally, the link traffic λ is updated according to the blocking probability to obtain the predicted performance index for the current iteration. A network performance prediction model is pre-trained using QT pseudo-labels and self-supervised labels obtained through contrastive learning.
[0120] Supervised learning module M3 for network performance prediction model: According to step S4, a certain amount of current network traffic features are collected in real time to perform supervised fine-tuning of the pre-trained network performance prediction model;
[0121] Dynamic network routing optimization module M4: includes the network performance prediction model, which adjusts the routing selection of forwarding devices based on the input network traffic characteristic data and the SLA of multiple candidate paths.
[0122] Specifically, firstly, for each origin node i and destination node j, select no more than k candidate paths to update the routing table matrix R, i.e., Ri i,j ={path1,…,path k The candidate paths are selected based on the network topology and the path length algorithm between two points i and j, with the path length less than the shortest length plus a threshold (thresh), and the candidate path R is guaranteed to be... i,j The total number is less than or equal to the threshold k.
[0123] Further, based on the network performance prediction model, the latency of the real-time network data flow under the given path in the current routing table is calculated. Combining the prediction results with the SLA threshold table, if the current traffic that meets the SLA increases, the iteration count is reset, and the current data flow is filtered and updated to Sat and No_Sat. After reaching the maximum optimization round, the routing table is completed and sent to the forwarding device to complete the routing optimization of the dynamic network.
Claims
1. A dynamic network routing optimization method based on deep learning prediction, characterized in that, include: Step S1: For the end-to-end network traffic input in the network topology, the path traversed by the network traffic, the routing buffer queue and multi-hop link information, and the corresponding statistical indicators are combined to finally construct a graph network with heterogeneous nodes. Step S2: Based on heterogeneous graph information, use the queuing theory QT algorithm to generate a large number of pseudo-labels for network traffic features regarding latency and packet loss metrics, including network performance indicators. Step S3: Construct a network performance prediction model and pre-train the model based on heterogeneous graph networks and network traffic feature data with pseudo-labels; Step S3 adopts the following: Step S3.1: Construct a network prediction model based on GNN message passing and attention. First, the path sequence with queue features and link features is combined to obtain the temporal representation of the path. Second, the path representation is used as the queue sequence to update the queue representation. Finally, the queue is used as the link sequence to update the link representation. The feature representation of the path node is obtained through multiple iterations. Step S3.2: Construct the loss function for the pre-trained model, including fitting the QT pseudo-labels from the performance metrics, denoted as... The regression / classification loss, flow-level contrastive loss, and network-level contrastive loss are compared with the model's predicted values. Step S3.3: Pre-train a network performance index model based on QT pseudo-label loss and same-sign comparison loss until convergence; Step S4: Based on the pre-trained network performance prediction model, fine-tuning training of the model is carried out using heterogeneous graph networks and a small amount of network traffic feature data with real labels. Step S4 adopts the following: Step S4.1: Similar to the network prediction model based on GNN message passing and attention in step S3.1, the model provides prediction results for the given network topology and various traffic characteristics; Step S4.2: Construct the loss function for model fine-tuning, i.e., fitting the true labels from the performance metrics, denoted as . The regression / classification loss of the model predictions, and the loss functions for latency and packet loss metrics can be expressed as follows; Step S4.3: Fine-tune the network performance index model based on the real label loss performance index until convergence; Step S5: Using the network prediction model trained in the above steps, compare different routing methods to obtain the performance indicators of each traffic flow in the current network, and select the better-performing solution as the optimized routing solution. The network performance prediction model is based on the GNN (Graph Transformer) model and MLP module in the pre-training stage, only the pre-trained GNN model is used in the fine-tuning stage, and the fine-tuned GNN model is used in the route optimization stage.
2. The dynamic network routing optimization method based on deep learning prediction according to claim 1, characterized in that, Step S1 adopts the following: Step S1.1: For the input network traffic, denote the terminal nodes in the network traffic as... End-to-end data stream, denoted as The buffer queue of the routing node and the one-hop link are denoted as follows: As nodes on the heterogeneous graph, and based on the links that the data flow passes through and the corresponding node relationships, edges are formed between heterogeneous nodes. Finally, the paths, links, and queue information of each data flow in the network topology are organized into a heterogeneous graph network. Step S1.2: Statistically record the traffic information of data packets in the network topology in multi-hop routing and global end-to-end according to a certain time interval, so as to represent the data packet throughput and bit rate statistics at the current moment, thereby forming the traffic characteristics of the heterogeneous graph network.
3. The dynamic network routing optimization method based on deep learning prediction according to claim 2, characterized in that, Step S1.2 adopts the following: Step S1.2.1: Record the routing and forwarding matrix of the current network topology, denoted as... ; Given a network data flow with a fixed routing table, record the traffic matrix. ,element Represents a node With nodes It is directly accessible; Step S1.2.2: Record the traffic matrix of the current network topology, denoted as... ; Considering the average bandwidth between each pair of nodes at the current moment, the total number of data packets, the distribution type of data packet arrival time and its distribution parameters, and the distribution type of data packet size and its distribution parameters, a dictionary structure is formed and stored in... middle; Step S1.2.3: Record the performance matrix of the current network topology, denoted as... ; The packet loss and delay quantile information between each pair of nodes at the current moment are used as... The element, the actual packet loss or latency indicator is marked as ; Step S1.2.4: Record the static characteristics of each heterogeneous node and the edge types between heterogeneous nodes; Record the characteristics of each node in the heterogeneous graph: Terminal nodes include the size of their buffer queues, representing the number of data packets stored, which can be denoted as... The link node includes the bandwidth of the two nodes and the scheduling policy characteristics used to serve the QoS queue. Queue nodes include buffer queue size and scheduling policy characteristics; The path nodes include scheduling policy type, data flow service type (ToS), packet arrival time distribution parameters, packet size distribution parameters, average bandwidth, average number of packets, latency, and packet loss characteristics. Record the characteristics of heterogeneous nodes, including the corresponding buffer queue size of the routing node and the static characteristics of the data flow path length; Record the edge types in the heterogeneous graph: the edge type between a queue node and a path or link node is marked as -1; the edge type between a path node and a link node is marked as 0; the edge type between a path node and a terminal node is marked as 1; and the edge type between a link node and a terminal node is marked as 2.
4. The dynamic network routing optimization method based on deep learning prediction according to claim 1, characterized in that, Step S2 adopts the following: Step S2.1: Obtain the network traffic routing relationship of each node at the current time step; In the heterogeneous graph, select edges of type 0 that consist of path nodes and link nodes. For each starting node, select the first set of edges, the second set of edges, and so on. Use these set of edges as different network traffic at each time step. Step S2.2: Based on the traffic characteristics of the link and path nodes in Step 1.2, initialize the average traffic for each path. Bandwidth capacity of each link Link traffic Link utilization ; Step S2.3: Calculate the utilization rate and blocking probability of each link; terminal node With terminal nodes The link utilization rate at the current time step can be calculated by dividing the cumulative traffic of the link by the link bandwidth capacity, expressed as: Further calculate the link blocking probability at the current time step, which is the probability of queuing delay occurring based on the buffer queue size, expressed as: ; Step S2.4: Update link traffic based on blocking probability And obtain the predicted packet loss for the current iteration; The link traffic is initialized to empty, and the first edge set is assigned the initial traffic value. The subsequent edge set is accumulated based on the blocking probability of each link, and the traffic of the current path in the next hop link is added. Then, the traffic is aggregated according to the corresponding edge index. After looping through multiple hop links, the final link traffic is obtained. This can be further expressed as follows; ； Step S2.5: Continuously update the blocking probability until there is no absolute difference in the traffic of a link before and after two updates that is greater than the threshold of 0.001; Step S2.6: Exit the loop and calculate the time delay prediction for each path node; First, calculate the probability that the link queue is empty, denoted as . , and its average utilization rate, are denoted as By aggregating the delays of all links on a given path, the delay of that path can be obtained, which represents the cumulative queuing delay along that path. 。 5. The dynamic network routing optimization method based on deep learning prediction according to claim 1, characterized in that, Step S3.1 adopts the following: Step S3.1.1: Initialize feature representation; The feature input of a given heterogeneous node is transformed into three feature representations of heterogeneous nodes. The path node features are initialized as average bandwidth and average number of packets, denoted as . The characteristics of a link node are the number of links, the link capacity, and the scheduling policy type, denoted as... The initialization of queue node characteristics includes queue size, scheduling policy, and weight, denoted as . ; Step S3.1.2: Update path features; Choosing edge types -1 and 0 in the heterogeneous graph represents the characteristics of link and queue nodes that are related to path nodes. Using a recurrent neural network (RNN) to learn these path features, it can be represented as follows: ,in This means concatenating the links and queues along the path according to the feature dimension and using them as the input sequence; This represents the initial hidden state in the RNN. The new path state representation is obtained through the RNN update and used for subsequent reading function performance metric prediction. Step S3.1.3: Update queue characteristics; Obtain the path state representation of the current iteration. Then, it is repartitioned based on edges of type 0 to obtain the input path sequence about the queue, and then... The RNN is updated using the initial hidden state; Step S3.1.4: Update link characteristics; Similarly, it is repartitioned based on edges of type -1 to obtain the input queue sequence for the links, and then... The RNN is updated using the initial hidden state; Step S3.1.5: Predict performance metrics; After several iterations, the state of each path is represented. A self-attention module is used for global representation, enabling contextual interaction between paths. Then, an MLP module with ReLU activation maps path features to performance metric prediction results, denoted as... .
6. The dynamic network routing optimization method based on deep learning prediction according to claim 1, characterized in that, Step S3.2 adopts the following: Step S3.2.1: Construct regression or classification losses for latency or packet loss metrics, using the MAPE loss function and the BCE loss function respectively. ； Among them This refers to the pseudo-labels of the performance metrics generated based on QT in step S2. This represents the model's performance metric prediction results for the current network data; Step S3.2.2: Construct the flow-level contrast loss; The input comes from two types of network datasets, either two network datasets from the same scenario or two network datasets from different scenarios, along with their network topologies and heterogeneous node features. Through step S3.1, performance prediction results for different data streams are obtained, denoted as... and Construct the flow-level contrast loss as ； The method uses an MLP with ReLU and sigmoid activation functions to convert the difference between the predictions of the two networks into floating-point numbers between 0 and 1. Step S3.2.3: Construct network-level contrastive loss; Similar to step S3.2.3, but averaging the difference results and constructing a network-level contrastive loss by fitting regression values. ； Step S3.2.4: Construct the overall loss of the pre-trained model; The model pre-training loss for latency performance metrics is: The model pre-training loss for packet loss performance metrics is: 。 7. The dynamic network routing optimization method based on deep learning prediction according to claim 1, characterized in that, Step S5 employs the following: Step S5.1: First, for each starting node... and destination node No more than 100 choices The candidate path set is used to update the routing table matrix. ,Right now The paths in the candidate set are determined by the network topology and the two points. and The path length algorithm selects paths whose length is less than the shortest length plus a threshold. And ensure candidate paths The total number is less than or equal to the threshold ; Step S5.2: Given an SLA threshold table for different data stream types and a set of traffic streams that meet the SLA, denoted as... The set of traffic that does not meet the SLA is denoted as... ; Step S5.3: Further calculate the real-time network data flow latency under the given path in the current routing table based on the network performance prediction model. Combine the prediction result with the SLA threshold table. If the traffic that meets the SLA increases, reset the iteration count, filter the current data flow, and update the table. middle; Step S5.4: From and Randomly select 1 / 5 of each traffic set, and randomly select one path from the candidate path set to form a new routing table for one iteration. Repeat step S5.3, and calculate a total of [number missing] paths. "times" indicates the number of iterations per round; Step S5.5: Specify the number of rounds, repeat step S5.4, and if the number of rounds exceeds the specified number, return the final routing table to guide the transmission of data streams.
8. A dynamic network routing optimization system based on deep learning prediction, characterized in that, Including the dynamic network routing optimization method based on deep learning prediction as described in any one of claims 1-7: Network feature acquisition module M1: Based on multiple routing and forwarding devices under a specified topology, it collects network traffic data at a certain time interval and, according to the characteristic requirements of heterogeneous nodes in step S1, statistically analyzes network features to form a network traffic feature sequence. Network performance prediction model pre-training module M2: Based on steps S2 and S3, pre-train the network performance prediction model using the QT algorithm and contrastive learning strategy; Supervised learning module M3 for network performance prediction model: According to step S4, a certain amount of current network traffic features are collected in real time to perform supervised fine-tuning of the pre-trained network performance prediction model; Dynamic network routing optimization module M4: includes the network performance prediction model, which adjusts the routing selection of forwarding devices based on the input network traffic characteristic data and the SLA of multiple candidate paths.