Dynamic routing optimization transmission method based on neural network

Abstract

The invention discloses a dynamic routing optimization transmission method based on a neural network. The method comprises the following steps: S1, collecting original state data in a network environment; s2, preprocessing the original state data; s3, constructing a causal graph based on the historical abnormal event log, and performing vectorization representation on the causal graph; s4, modeling is carried out on the state features and the topological features, state and structure fusion processing is carried out, and a performance prediction tensor is output through a multi-layer perceptron; s5, constructing a joint optimization target based on the embedded vector and the performance prediction tensor, and screening an optimal path set; and S6, generating a routing control instruction according to the optimal path set, and issuing the routing control instruction to the network forwarding equipment. According to the method, causal graph embedding and neural network modeling are fused, dynamic path optimization is realized, and the method has the advantages of high intelligence, high stability and high adaptability.

Description

Technical Field

[0001] This invention relates to the field of network communication, and in particular to a dynamic routing optimization transmission method based on neural networks. Background Technology

[0002] In existing network communication systems, dynamic routing algorithms are the core mechanism for ensuring data transmission efficiency and network stability. Traditional dynamic routing protocols such as RIP and OSPF mainly rely on link state information and static cost functions for path selection. Although these protocols have a certain degree of network adaptability, their decision-making process is usually based on local information or simple heuristic rules, which cannot effectively capture the high-dimensional feature interaction relationships in network state changes. At the same time, when faced with sudden events in complex network environments such as link congestion, node failure, or path anomalies, traditional algorithms are slow to respond and have poor adaptability, which can easily lead to routing oscillations, increased data packet loss rate, and decreased network resource utilization.

[0003] With the development of artificial intelligence, more and more research attempts to introduce deep learning technologies such as neural networks into routing decisions, hoping to leverage their advantages in feature representation and nonlinear modeling to achieve more intelligent path optimization. Some methods use graph neural networks to model the network topology and extract global features by aggregating node information, but they often only focus on structural modeling and ignore the causal relationship between time series information and historical behavioral anomalies. Other studies use reinforcement learning or predictive models to model the network state, but most of these methods lack explicit modeling mechanisms for the evolution of anomalies, making it difficult to effectively improve the generalization ability and path robustness in complex and abnormal environments. In addition, existing AI routing optimization methods generally face problems such as poor interpretability of path selection, weak anomaly tolerance, and insufficient path controllability.

[0004] Based on the above problems, how to integrate high-dimensional network state data, historical abnormal behavior causal relationships and topological characteristics to construct a learnable, reasonable and updatable dynamic routing optimization mechanism has become a key breakthrough that current intelligent network scheduling technology urgently needs to make. Summary of the Invention

[0005] One objective of this invention is to propose a dynamic routing optimization transmission method based on neural networks. This invention integrates causal graph embedding and structural state dual decoupled neural network modeling methods, constructs a causal relationship graph based on historical abnormal events, and realizes intelligent screening and dynamic optimization of candidate paths through graph embedding and path scoring mechanisms. It has the advantages of high path selection accuracy, strong network adaptability, and strong anomaly tolerance, and is suitable for complex network environments with high dynamics and high reliability requirements.

[0006] A dynamic routing optimization transmission method based on a neural network according to an embodiment of the present invention includes the following steps: S1. Collect raw state data from the network environment; S2. Preprocess the original state data to generate the network state tensor; S3. Construct a cause-effect graph based on historical abnormal event logs, and use a graph embedding algorithm to vectorize the cause-effect graph to generate an embedding vector. S4. Input the network state tensor and topological adjacency matrix into the structure-state dual decoupled neural network, model the state features and topological features respectively, perform state and structure fusion processing, and output the performance prediction tensor through a multilayer perceptron. S5. Construct a joint optimization objective based on the embedding vector and the performance prediction tensor, and select the optimal path set; S6. Generate routing control instructions based on the optimal path set and send them to the network forwarding devices.

[0007] Optionally, the raw state data includes link latency, bandwidth utilization, packet loss rate, path hop count, CPU utilization, and topology adjacency matrix. The topology adjacency matrix indicates whether there is a valid link connection between network nodes at a certain moment. If there is, the element value of the topology adjacency matrix is ​​1; otherwise, it is 0.

[0008] Optionally, the preprocessing includes field alignment, timestamp unification, data cleaning, missing value imputation, and normalization.

[0009] Optionally, S3 specifically includes: S31. Extract historical abnormal event logs and construct a raw event record set. Each record is represented as a triple. , Indicates the time when the event occurred. This represents the set of nodes that are affected. Indicates the category label to which the event belongs. This represents the total number of historical events. S32. Construct an event pair set based on the time series and tag sequence of events in the original event record set. The joint occurrence frequency, conditional probability, and average interval time of each pair of events are statistically analyzed. The joint occurrence frequency represents the number of times two event types occur together, and the conditional probability represents the conditional probability of another event occurring after one event occurs. S33. Using the category label of the event as nodes, construct a node set; based on the conditional dependencies of the event on the set, establish an edge set; and define the cause-effect graph structure as follows. ,in This represents a directed causal graph. Represents a set of nodes. The set of edges is defined by determining whether the mutual information between two events is greater than a threshold. If it is greater, then an edge is established. S34. For each edge in the edge set, calculate the causal weight: ; in, Represents a node To the node Causal weights The weighting coefficients of the conditional probability factors are represented. The weighting coefficients of the frequency factor. The weighting coefficients representing the time stability factor. Represents conditional probability. Indicates an event and The frequency of their joint occurrence Indicates the standardized joint frequency. Indicates an event and The average time interval, This represents the average time interval between all event pairs. Indicates the first One event, Indicates the first One event; S35. The causal graph with causal weights is input into the graph embedding algorithm and vectorized to generate an embedding vector. The graph embedding algorithm uses the DeepWalk algorithm with causal constraints.

[0010] Optionally, S35 specifically includes: S351, Perform the following for each node in the cause-effect graph node set: The second length is Random walks generate a set of node paths. ,in This indicates the total number of paths. Indicates the first A walking sequence, Indicates the th step in the walking sequence 1 node Indicates a random length; S352. During the traversal, the jumps between nodes are controlled according to the causal weights of the causal edges, defining the jump from the current node... to adjacent nodes Jump probability: ; in, Indicates from node Jump to node The jump probability, Represents a node The set of adjacent nodes, Represents a node To the node Causal weights Indicates "belongs to"; S353. The set of node paths is used as training corpus and input into the improved Skip-Gram model for embedding learning training. Embedding vectors of nodes are generated through context training. The training process includes: Define the causal density and calculate the context window radius based on the causal density of each node in the causal graph: ; in, Represents a node causal density, Represents a node To the node Causal weights; ; in, Represents a node The radius of the context window, Indicates the window adjustment factor. This indicates a round-down operation. This represents the maximum limit value of the context window radius. This indicates taking the minimum value; For each node in each path, extract context nodes that do not exceed the radius of the context window. Construct sample pairs based on the center node and the context nodes. Each sample pair constitutes a training sample. Merge the sample pairs of all paths to obtain a sample pair set. Adjusting the number of negative samples based on causality density: ; in, Represents a node The number of negative samples, This represents the negative sampling adjustment factor; For each pair of samples, randomly sample from the set of sample pairs. The negative sample set consists of nodes that do not form a sample pair with the central node; For each sample pair and its corresponding negative sample node Construct the optimization objective function: ; in, This represents the objective function to be optimized. Represents a node Embedded vector, Represents a node Embedded vector, Represents negative sample nodes Embedded vector, Representing vectors and The inner product, Represents a node Embedded vector, This represents the Sigmoid activation function. Represents the set of sample pairs. Represents the set of negative samples. Indicates transpose; The embedding vectors of nodes are trained using stochastic gradient descent, iterating until the objective function converges or the training epochs are reached, and the embedding vector of each node after training is output.

[0011] Optionally, S4 specifically includes: S41. Input the network state tensor and topological adjacency matrix into the structure-state dual decoupled neural network, the structure-state dual decoupled neural network including a state subnetwork and a structure subnetwork, the state subnetwork using a gated recurrent unit network, and the structure subnetwork using a graph convolutional network; S42. Input the network state tensor into the state sub-network to generate the state embedding tensor; S43. Input the network state tensor and the topological adjacency matrix into the structural subnetwork, and extract the structural feature tensor through graph convolution operation: ; in, Represents the structural feature tensor. Indicates the graph convolution parameters. Represents the topological adjacency matrix; S44. The state embedding tensor and the structural feature tensor are concatenated along the channel dimension to form a fused feature tensor, which is then input into a multilayer perceptron to obtain a performance prediction tensor. The performance prediction tensor includes predicted latency, load, and stability scores.

[0012] Optionally, S5 specifically includes: S51. Initialize attention scoring vectors randomly using a Gaussian distribution, and generate path embedding representations based on node embedding vectors using an attention mechanism: ; in, This represents path embedding representation. Represents a node Attention weights This represents the attention scoring vector. Represents the attention weight matrix. Indicates the first in the path The embedding vector of each node. Indicates the first in the path The embedding vector of each node. This represents the natural exponential function. This represents the hyperbolic tangent activation function. Indicates transpose; S52. Extract the latency, load, and stability scores for each path from the performance prediction tensor, and calculate the path's performance score accordingly: ; in, Representing a path Performance rating Indicates the weight of the latency metric. The weights of the load metrics are indicated. The weights of the stability index are represented. Representing a path Latency score, Representing a path Load score, Representing a path Stability score, Represents the paths in the set of node paths; S53. For each path, calculate the Euclidean distance between the path embedding representation and the normal path reference embedding representation as the semantic deviation. The normal path reference embedding representation is obtained by calculating the average value of the embedding vectors of all nodes. S54. Perform causal intervention simulation on the node set of each path. For any two nodes in the node set, calculate the hop count of the causal path using a breadth-first search algorithm, and obtain the causal sensitivity score of the path: ; in, Representing a path Causal sensitivity score Representing a path The number of corresponding nodes Indicates from node To the node Causal weights Represents a node arrive Path jump count of causal path The set of nodes representing the path; S55. Construct a joint optimization objective based on performance score, semantic deviation, and causal sensitivity score, and select the objective with the smallest joint optimization objective. The set of paths is considered as the optimal path set. ; in, Represents the set of optimal paths. An adjustment coefficient representing the degree of semantic deviation. The moderating coefficient representing the causal sensitivity score. Indicates semantic deviation. This means selecting the option with the smallest objective function. A path.

[0013] Optionally, S6 specifically includes: S61. Receive the set of optimal paths output, wherein each path in the set of optimal paths consists of several network nodes arranged in sequence; S62. Based on the order of nodes in each path, extract the forwarding relationship of each hop, and form a forwarding table entry set consisting of source node, destination node and next hop node to describe the hop-by-hop forwarding structure of the path. S63. Generate corresponding control instructions based on the forwarding table entry set, wherein the control instructions are encapsulated in a format supported by the network controller; S64. The control commands are grouped according to the target devices and sent to the corresponding network forwarding devices through the controller interface; S65. After receiving the control command, the network device configures the forwarding behavior according to the command content; S66. After the controller completes the instruction, it updates the path status information and records the activation status of each path and the network node information it covers.

[0014] The beneficial effects of this invention are: First, by introducing causal modeling and graph embedding learning mechanisms, this invention improves the intelligence level and responsiveness of dynamic routing optimization in complex network environments. By constructing a causal graph based on historical abnormal event logs and using a graph embedding algorithm with causal constraints to learn the representation of event nodes, this invention can not only characterize the deep causal relationships in network state evolution, but also achieve high-dimensional modeling of path semantics.

[0015] Secondly, in the routing decision-making process, the structural-state dual-decoupled neural network designed in this invention models network state features and topological features separately, and expresses them uniformly through a fusion mechanism, achieving efficient integration and deep representation of heterogeneous data in dynamic networks. Combined with the performance prediction module of a multilayer perceptron, it can perform multi-dimensional quantitative prediction of the latency, load, and stability of candidate paths, further improving the accuracy and foresight of the routing strategy. By constructing a joint optimization objective, it comprehensively evaluates path performance, semantic deviation, and causal sensitivity, ultimately selecting the optimal set of paths, effectively improving the system's adaptability to complex scenarios such as sudden anomalies and link disturbances.

[0016] In addition, this invention also designs a path-level causal intervention and semantic analysis module, which can introduce intelligent reasoning at the network behavior level during the routing process, realize refined and personalized control of path selection, and finally achieve efficient command issuance to forwarding devices through standardized control commands, ensuring the smoothness of the path switching process and the control loop. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0018] Figure 1 This is a flowchart of a dynamic routing optimization transmission method based on neural networks proposed in this invention; Figure 2 This is a schematic diagram of a graph embedding learning structure based on a causal graph, which is a dynamic routing optimization transmission method based on a neural network proposed in this invention. Figure 3 This is a diagram showing the generation of the performance prediction tensor and the path scoring structure of a dynamic routing optimization transmission method based on neural networks proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figure 1-3 A dynamic routing optimization transmission method based on neural networks includes the following steps: S1. Collect raw state data from the network environment; S2. Preprocess the original state data to generate the network state tensor; S3. Construct a cause-effect graph based on historical abnormal event logs, and use a graph embedding algorithm to vectorize the cause-effect graph to generate an embedding vector. S4. Input the network state tensor and topological adjacency matrix into the structure-state dual decoupled neural network, model the state features and topological features respectively, perform state and structure fusion processing, and output the performance prediction tensor through a multilayer perceptron. S5. Construct a joint optimization objective based on the embedding vector and the performance prediction tensor, and select the optimal path set; S6. Generate routing control instructions based on the optimal path set and send them to the network forwarding devices.

[0021] This invention proposes a dynamic routing optimization transmission method based on neural networks, which integrates multiple core technologies such as state awareness, causal modeling, and structural decoupling learning. It effectively improves the adaptability of routing decisions to complex network state changes. By introducing a joint mechanism of performance prediction and causal reasoning, it can intelligently select the optimal path, reduce the risk of routing oscillation, and improve the overall stability and throughput of the system. It is suitable for intelligent network scheduling needs in high-concurrency and highly dynamic scenarios, and is significantly superior to traditional heuristic or static rule-based path selection methods.

[0022] In this embodiment, the original state data includes link latency, bandwidth utilization, packet loss rate, path hop count, CPU utilization, and topology adjacency matrix. The topology adjacency matrix indicates whether there is a valid link connection between network nodes at a certain time. If there is, the element value of the topology adjacency matrix is ​​1, otherwise it is 0.

[0023] This invention collects multi-dimensional raw state data such as link latency, bandwidth utilization, packet loss rate, path hop count, and CPU utilization, and combines this with the structural information of the topology adjacency matrix to provide a comprehensive and accurate input foundation for subsequent network state modeling and path optimization. This multi-indicator fusion collection strategy enhances the system's ability to identify hidden problems such as network performance bottlenecks and abnormal node loads, and improves the stability and robustness of data-driven modeling. It is an indispensable prerequisite for achieving deep intelligent optimization.

[0024] In this embodiment, the preprocessing includes field alignment, timestamp unification, data cleaning, missing value imputation, and normalization.

[0025] In the process of preprocessing the original network state data, this invention fully considers the temporality, consistency and integrity of the data. Through field alignment, timestamp unification, data cleaning, missing value imputation and normalization, it effectively eliminates redundant and noisy information, improves the training convergence speed and expressive power of subsequent neural network modeling, and the high-quality preprocessing results construct a network state tensor with stable structure and clear semantics, providing a solid data foundation for performance prediction and path scoring.

[0026] In this embodiment, S3 specifically includes: S31. Extract historical abnormal event logs and construct a raw event record set. Each record is represented as a triple. , Indicates the time when the event occurred. This represents the set of nodes that are affected. Indicates the category label to which the event belongs. This represents the total number of historical events. S32. Construct an event pair set based on the time series and tag sequence of events in the original event record set. The joint occurrence frequency, conditional probability, and average interval time of each pair of events are statistically analyzed. The joint occurrence frequency represents the number of times two event types occur together, and the conditional probability represents the conditional probability of another event occurring after one event occurs. S33. Using the category label of the event as nodes, construct a node set; based on the conditional dependencies of the event on the set, establish an edge set; and define the cause-effect graph structure as follows. ,in This represents a directed causal graph. Represents a set of nodes. The set of edges is defined by determining whether the mutual information between two events is greater than a threshold. If it is greater, then an edge is established. S34. For each edge in the edge set, calculate the causal weight: ; in, Represents a node To the node Causal weights The weighting coefficients of the conditional probability factors are represented. The weighting coefficients of the frequency factor. The weighting coefficients representing the time stability factor. Represents conditional probability. Indicates an event and The frequency of their joint occurrence Indicates the standardized joint frequency. Indicates an event and The average time interval, This represents the average time interval between all event pairs. Indicates the first One event, Indicates the first One event; S35. The causal graph with causal weights is input into the graph embedding algorithm and vectorized to generate an embedding vector. The graph embedding algorithm uses the DeepWalk algorithm with causal constraints.

[0027] This invention effectively captures the potential causal relationships between events in network operation by constructing a causal graph based on historical abnormal event logs. It breaks through the limitations of traditional methods that rely solely on time series modeling. By introducing causal weight calculation and graph embedding representation, it can characterize the long-term dependencies and intervention sensitivity between events, providing causal explanation and reasoning support for path optimization, enhancing the interpretability and predictive accuracy of network behavior modeling, and improving the ability to identify and avoid abnormal paths.

[0028] In this embodiment, S35 specifically includes: S351, Perform the following for each node in the cause-effect graph node set: The second length is Random walks generate a set of node paths. ,in This indicates the total number of paths. Indicates the first A walking sequence, Indicates the th step in the walking sequence 1 node Indicates a random length; S352. During the traversal, the jumps between nodes are controlled according to the causal weights of the causal edges, defining the jump from the current node... to adjacent nodes Jump probability: ; in, Indicates from node Jump to node The jump probability, Represents a node The set of adjacent nodes, Represents a node To the node Causal weights Indicates "belongs to"; S353. The set of node paths is used as training corpus and input into the improved Skip-Gram model for embedding learning training. Embedding vectors of nodes are generated through context training. The training process includes: Define the causal density and calculate the context window radius based on the causal density of each node in the causal graph: ; in, Represents a node causal density, Represents a node To the node Causal weights; ; in, Represents a node The radius of the context window, Indicates the window adjustment factor. This indicates a round-down operation. This represents the maximum limit value of the context window radius. This indicates taking the minimum value; For each node in each path, extract context nodes that do not exceed the radius of the context window. Construct sample pairs based on the center node and the context nodes. Each sample pair constitutes a training sample. Merge the sample pairs of all paths to obtain a sample pair set. Adjusting the number of negative samples based on causality density: ; in, Represents a node The number of negative samples, This represents the negative sampling adjustment factor; For each pair of samples, randomly sample from the set of sample pairs. The negative sample set consists of nodes that do not form a sample pair with the central node; For each sample pair and its corresponding negative sample node Construct the optimization objective function: ; in, This represents the objective function to be optimized. Represents a node Embedded vector, Represents a node Embedded vector, Represents negative sample nodes Embedded vector, Representing vectors and The inner product, Represents a node Embedded vector, This represents the Sigmoid activation function. Represents the set of sample pairs. Represents the set of negative samples. Indicates transpose; The embedding vectors of nodes are trained using stochastic gradient descent, iterating until the objective function converges or the training epochs are reached, and the embedding vector of each node after training is output.

[0029] This invention introduces the DeepWalk algorithm with causal constraints into the causal graph embedding process, and combines it with a context window dynamically adjusted by causal density and a negative sampling strategy to achieve high-quality learning of node embeddings, thereby improving the semantic representation capability of the graph. By constructing training corpora through sampling paths, the model is guided to pay more attention to the key node relationships in the causal path, and the resulting embedding vectors have stronger discriminative power and generalization ability, laying an important foundation for the construction of path semantic scoring and optimization objective function.

[0030] In this embodiment, S4 specifically includes: S41. Input the network state tensor and topological adjacency matrix into the structure-state dual decoupled neural network, the structure-state dual decoupled neural network including a state subnetwork and a structure subnetwork, the state subnetwork using a gated recurrent unit network, and the structure subnetwork using a graph convolutional network; S42. Input the network state tensor into the state sub-network to generate the state embedding tensor; S43. Input the network state tensor and the topological adjacency matrix into the structural subnetwork, and extract the structural feature tensor through graph convolution operation: ; in, Represents the structural feature tensor. Indicates the graph convolution parameters. Represents the topological adjacency matrix; S44. The state embedding tensor and the structural feature tensor are concatenated along the channel dimension to form a fused feature tensor, which is then input into a multilayer perceptron to obtain a performance prediction tensor. The performance prediction tensor includes predicted latency, load, and stability scores.

[0031] This invention employs a structure-state dual-decoupled neural network to model the network state tensor and topological adjacency matrix in parallel, extracting state features and structural features respectively, and then fusing them into a unified representation before inputting it into a multilayer perceptron for performance prediction. This structure avoids mutual interference between state and structural information, while improving the model's adaptability to dynamically changing network structures. It can more accurately predict the latency, load, and stability of candidate paths in the future time period, achieving more precise path quality assessment.

[0032] In this embodiment, S5 specifically includes: S51. Initialize attention scoring vectors randomly using a Gaussian distribution, and generate path embedding representations based on node embedding vectors using an attention mechanism: ; in, This represents path embedding representation. Represents a node Attention weights This represents the attention scoring vector. Represents the attention weight matrix. Indicates the first in the path The embedding vector of each node. Indicates the first in the path The embedding vector of each node. This represents the natural exponential function. This represents the hyperbolic tangent activation function. Indicates transpose; S52. Extract the latency, load, and stability scores for each path from the performance prediction tensor, and calculate the path's performance score accordingly: ; in, Representing a path Performance rating Indicates the weight of the latency metric. The weights of the load metrics are indicated. The weights of the stability index are represented. Representing a path Latency score, Representing a path Load score, Representing a path Stability score, Represents the paths in the set of node paths; S53. For each path, calculate the Euclidean distance between the path embedding representation and the normal path reference embedding representation as the semantic deviation. The normal path reference embedding representation is obtained by calculating the average value of the embedding vectors of all nodes. S54. Perform causal intervention simulation on the node set of each path. For any two nodes in the node set, calculate the hop count of the causal path using a breadth-first search algorithm, and obtain the causal sensitivity score of the path: ; in, Representing a path Causal sensitivity score Representing a path The number of corresponding nodes Indicates from node To the node Causal weights Represents a node arrive Path jump count of causal path The set of nodes representing the path; S55. Construct a joint optimization objective based on performance score, semantic deviation, and causal sensitivity score, and select the objective with the smallest joint optimization objective. The set of paths is considered as the optimal path set. ; in, Represents the set of optimal paths. An adjustment coefficient representing the degree of semantic deviation. The moderating coefficient representing the causal sensitivity score. Indicates semantic deviation. This means selecting the option with the smallest objective function. A path.

[0033] This invention introduces a joint optimization mechanism combining path semantic scoring, graph semantic deviation, and causal sensitivity in the optimal path selection stage, breaking through the traditional single-dimensional decision-making model of path selection based solely on performance prediction. Through the weighted integration of multi-dimensional factors, it can identify potential abnormal paths, improve the robustness and security of path selection, and maintain stable decision output, especially when the network experiences sudden state changes, demonstrating strong generalization ability and intelligent intervention level.

[0034] In this embodiment, S6 specifically includes: S61. Receive the set of optimal paths output, wherein each path in the set of optimal paths consists of several network nodes arranged in sequence; S62. Based on the order of nodes in each path, extract the forwarding relationship of each hop, and form a forwarding table entry set consisting of source node, destination node and next hop node to describe the hop-by-hop forwarding structure of the path. S63. Generate corresponding control instructions based on the forwarding table entry set, wherein the control instructions are encapsulated in a format supported by the network controller; S64. The control commands are grouped according to the target devices and sent to the corresponding network forwarding devices through the controller interface; S65. After receiving the control command, the network device configures the forwarding behavior according to the command content; S66. After the controller completes the instruction, it updates the path status information and records the activation status of each path and the network node information it covers.

[0035] This invention refines the entire process of extracting hop-by-hop structure from the optimal path set, constructing forwarding table entries, generating instructions, distributing packets, and configuring network devices in the routing control instruction issuance process. This ensures the final controllability and execution loop of the path optimization scheme. The method ensures a complete mapping of the routing policy from the logical layer to the physical layer, avoids ambiguity or deviation in instruction execution, improves the transmission efficiency of controller instructions and device compatibility, and enhances the overall operability and reliability.

[0036] Example 1: To verify the feasibility of this invention in practice, it was applied to a network communication scheduling scenario in an industrial internet experimental park. The park has deployed more than 80 industrial intelligent terminals, forming a heterogeneous network environment with high density and strong real-time requirements. The network structure is complex and the links fluctuate frequently. The network has several historical problems during operation, including: large fluctuations in data transmission path latency, untimely response to sudden link failures, and lack of interpretability in path selection. Traditional OSPF routing protocols and simple heuristic path planning methods are difficult to meet the requirements of intelligent, stable, and efficient scheduling.

[0037] To this end, the proposed neural network-based dynamic routing optimization transmission method is integrated into the network control system. First, the system automatically collects indicators such as link status data, bandwidth utilization, packet loss rate, CPU load, and topology connection information between nodes in the campus network to construct the original state tensor. At the same time, historical abnormal event logs are structured and causal graphs are constructed. Based on this, a graph embedding algorithm with causal constraints is applied to generate graph embedding vectors, providing strong causal semantic support for path behavior modeling.

[0038] Subsequently, the tensors are input into a structure-state dual-decoupled neural network. The state subnetwork uses a gated recurrent network (GRU) to capture state sequence features, while the structure subnetwork uses graph convolution to extract topological patterns. After fusing the state and structure representations, a multilayer perceptron is used to generate candidate path performance prediction results, including latency, load, and stability scores. In the path selection stage, the system constructs a multidimensional joint scoring model based on prediction tensors, graph embedding, and causal inference modules. Finally, the system selects the optimal path cluster and dynamically issues routing control commands. The comparative analysis results of the system with traditional dynamic routing methods after 30 days of operation are shown in Table 1.

[0039] Table 1 Comparison of Dynamic Routing Optimization Experiments

[0040] In terms of average path latency, the method of this invention is significantly better than the traditional method, with the average latency decreasing from 56.7ms to 41.4ms, a reduction of about 27%. This means that in high-density terminal deployment environments, the method of this invention can complete the data packet transmission task more quickly, thereby improving the overall communication efficiency. At the same time, the standard deviation of path latency decreased from 14.3ms to 8.7ms, indicating that the latency volatility decreased, the network performance was more stable, and it was more friendly to tasks with low jitter requirements.

[0041] Regarding the average packet loss rate, the present invention controls the packet loss rate to 1.6%, which is almost half of the 3.2% of the traditional method. The reduction in packet loss rate improves the reliability of data transmission, especially in congested or unstable link scenarios, where the present invention exhibits better robustness.

[0042] The path stability score improved from 0.58 to 0.79, indicating that the method of the present invention can maintain more consistent and predictable path performance, which is beneficial to supporting business systems with high real-time requirements such as industrial control and telemedicine. Furthermore, in terms of abnormal path identification rate, the present invention reached 89.3%, which is much higher than the 62.5% of the traditional method. This identification ability comes from the combination of causal graph modeling and graph embedding mechanism, which effectively improves the ability to perceive potentially high-risk paths.

[0043] In terms of emergency response capabilities, the path reconfiguration response time is reduced to 2.7 seconds, while the traditional method requires 8.1 seconds to complete path switching. This means that the present invention can adapt to network topology changes more quickly, improve network recovery speed, and reduce the risk of outages.

[0044] The success rate of control command issuance also increased from 92.6% to 98.4%, reflecting the improved reliability of the method of the present invention in command transmission between the controller and the forwarding device. Similarly, the average load balancing degree increased from 0.65 to 0.82, indicating that the use of network resources is more balanced, avoiding link bottlenecks and local congestion.

[0045] Finally, the node processing efficiency was improved by 17.8%, demonstrating the optimization benefits brought by this invention at the control system computation level, and providing stronger computing power support and parallel capabilities for path calculation and deployment strategies.

[0046] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A dynamic routing optimization transmission method based on neural networks, characterized in that, Includes the following steps: S1. Collect raw state data from the network environment; S2. Preprocess the original state data to generate the network state tensor; S3. Construct a cause-effect graph based on historical abnormal event logs, and use a graph embedding algorithm to vectorize the cause-effect graph to generate an embedding vector. S4. Input the network state tensor and topological adjacency matrix into the structure-state dual decoupled neural network, model the state features and topological features respectively, perform state and structure fusion processing, and output the performance prediction tensor through a multilayer perceptron. S5. Construct a joint optimization objective based on the embedding vector and the performance prediction tensor, and select the optimal path set; S6. Generate routing control instructions based on the optimal path set and send them to the network forwarding devices.

2. The dynamic routing optimization transmission method based on neural networks according to claim 1, characterized in that, The raw state data includes link latency, bandwidth utilization, packet loss rate, path hop count, CPU utilization, and topology adjacency matrix. The topology adjacency matrix indicates whether there is a valid link connection between network nodes at a certain moment. If there is, the element value of the topology adjacency matrix is ​​1, otherwise it is 0.

3. The dynamic routing optimization transmission method based on neural networks according to claim 1, characterized in that, The preprocessing includes field alignment, timestamp unification, data cleaning, missing value imputation, and normalization.

4. The dynamic routing optimization transmission method based on neural networks according to claim 1, characterized in that, S3 specifically includes: S31. Extract historical abnormal event logs and construct a raw event record set. Each record is represented as a triple. , Indicates the time when the event occurred. This represents the set of nodes that are affected. Indicates the category label to which the event belongs. This represents the total number of historical events. S32. Construct an event pair set based on the time series and tag sequence of events in the original event record set. Calculate the joint occurrence frequency, conditional probability, and average interval for each pair of events; S33. Using the category label of the event as the node, construct a node set, establish an edge set according to the conditional dependencies of the event on the set, and define the cause-effect graph structure; S34. For each edge in the edge set, calculate the causal weight: ； in, Represents a node To the node Causal weights The weighting coefficients of the conditional probability factors are represented. The weighting coefficients of the frequency factor. The weighting coefficients representing the time stability factor. Represents conditional probability. Indicates an event and The frequency of their joint occurrence Indicates the standardized joint frequency. Indicates an event and The average time interval, This represents the average time interval between all event pairs. Indicates the first One event, Indicates the first One event; S35. The causal graph with causal weights is input into the graph embedding algorithm and vectorized to generate an embedding vector. The graph embedding algorithm uses the DeepWalk algorithm with causal constraints.

5. The dynamic routing optimization transmission method based on neural networks according to claim 4, characterized in that, Specifically, S35 includes: S351, Perform the following for each node in the cause-effect graph node set: The second length is A random walk generates a set of node paths; S352. During the traversal, the jumps between nodes are controlled according to the causal weights of the causal edges, defining the jump from the current node... to adjacent nodes The jump probability; S353. The set of node paths is used as training corpus and input into the improved Skip-Gram model for embedding learning training. Embedding vectors of nodes are generated through context training. The training process includes: Define the causal density and calculate the context window radius based on the causal density of each node in the causal graph; For each node in each path, extract context nodes that do not exceed the radius of the context window. Construct sample pairs based on the center node and the context nodes. Each sample pair constitutes a training sample. Merge the sample pairs of all paths to obtain a sample pair set. Adjusting the number of negative samples based on causality density: ； in, Represents a node The number of negative samples, This represents the negative sampling adjustment factor; For each pair of samples, randomly sample from the set of sample pairs. The negative sample set consists of nodes that do not form a sample pair with the central node; For each sample pair and its corresponding negative sample node Construct the optimization objective function: ； in, This represents the objective function to be optimized. Represents a node Embedded vector, Represents a node Embedded vector, Represents negative sample nodes Embedded vector, Representing vectors and The inner product, Represents a node Embedded vector, This represents the Sigmoid activation function. Represents the set of sample pairs. Represents the set of negative samples. Indicates transpose; The embedding vectors of nodes are trained using stochastic gradient descent, iterating until the objective function converges or the training epochs are reached, and the embedding vector of each node after training is output.

6. The dynamic routing optimization transmission method based on neural networks according to claim 1, characterized in that, S4 specifically includes: S41. Input the network state tensor and topological adjacency matrix into the structure-state dual decoupled neural network, the structure-state dual decoupled neural network including a state subnetwork and a structure subnetwork, the state subnetwork using a gated recurrent unit network, and the structure subnetwork using a graph convolutional network; S42. Input the network state tensor into the state sub-network to generate the state embedding tensor; S43. Input the network state tensor and the topological adjacency matrix into the structural sub-network, and extract the structural feature tensor through graph convolution operation; S44. The state embedding tensor and the structural feature tensor are concatenated along the channel dimension to form a fused feature tensor, which is then input into a multilayer perceptron to obtain a performance prediction tensor. The performance prediction tensor includes predicted latency, load, and stability scores.

7. The dynamic routing optimization transmission method based on neural networks according to claim 1, characterized in that, S5 specifically includes: S51. Initialize the attention scoring vector randomly using a Gaussian distribution, and generate the path embedding representation based on the node's embedding vector using an attention mechanism; S52. Extract the latency, load, and stability scores for each path from the performance prediction tensor, and calculate the performance score of the path accordingly. S53. For each path, calculate the Euclidean distance between the path embedding representation and the normal path baseline embedding representation as the semantic deviation. S54. Perform causal intervention simulation on the node set of each path. For any two nodes in the node set, calculate the hop count of the causal path using a breadth-first search algorithm, and obtain the causal sensitivity score of the path: ； in, Representing a path Causal sensitivity score Representing a path The number of corresponding nodes Indicates from node To the node Causal weights Represents a node arrive Path jump count of causal path The set of nodes representing the path; S55. Construct a joint optimization objective based on performance score, semantic deviation, and causal sensitivity score, and select the objective with the smallest joint optimization objective. The path is taken as the set of optimal paths.

8. The dynamic routing optimization transmission method based on neural networks according to claim 1, characterized in that, S6 specifically includes: S61. Receive the set of optimal paths output, wherein each path in the set of optimal paths consists of several network nodes arranged in sequence; S62. Based on the order of nodes in each path, extract the forwarding relationship of each hop, and form a forwarding table entry set consisting of source node, destination node and next hop node to describe the hop-by-hop forwarding structure of the path. S63. Generate corresponding control instructions based on the forwarding table entry set, wherein the control instructions are encapsulated in a format supported by the network controller; S64. The control commands are grouped according to the target devices and sent to the corresponding network forwarding devices through the controller interface; S65. After receiving the control command, the network device configures the forwarding behavior according to the command content; S66. After the controller completes the instruction, it updates the path status information and records the activation status of each path and the network node information it covers.

Images