A Time-Sensitive Network Traffic Scheduling System and Method Based on Causal Data Completion

Abstract

This invention belongs to the field of communication and network control technology. To address the problems of poor adaptability in non-ideal observation environments and the inability of a single optimization objective to meet various differentiated service requirements, and to overcome scheduling failures caused by network jitter or node failures leading to data loss, this invention provides: a time-sensitive network traffic scheduling method based on causal data completion. This method collects real-time traffic data from the current time-sensitive network (TSN) and constructs an irregular observation matrix; it uses a state reconstruction module based on neural Granger causality to complete the missing data in the observation matrix; it extracts robust spatiotemporal features with noise resistance from the completed observation matrix using a variational graph attention network; and it inputs these robust spatiotemporal features into a reinforcement learning decision network based on hierarchical optimization. Under the hard real-time constraint of priority critical time-triggered TT flow, it optimizes the best-effort BE flow transmission efficiency, ultimately generating a gated control list for execution. This invention is primarily applied in design and manufacturing applications.

Description

Technical Field

[0001] This invention belongs to the field of communication and network control technology, and specifically relates to a traffic scheduling method for Time-Sensitive Networking (TSN) applied to communication and next-generation computing architectures, namely, a TSN traffic scheduling method based on causal data completion. More specifically, this invention relates to an intelligent and robust network resource scheduling scheme that integrates causal inference theory with graph neural networks (GNN) and deep reinforcement learning (DRL) techniques. Background Technology

[0002] The current electronic and electrical architecture is undergoing profound changes, with traffic scheduling architecture shifting from traditional distributed architectures to domain-centralized and even centralized computing architectures. Among these new architectures, Time-Sensitive Networking (TSN), with its ability to provide deterministic, low-latency, and highly reliable data transmission, has become the next-generation communication backbone connecting core components such as LiDAR, cameras, and domain controllers. The design of the TSN traffic scheduling mechanism—that is, how to rationally allocate network time slot resources to ensure the quality of service for different types of data streams—is directly related to network reliability.

[0003] Currently, research on TSN traffic scheduling mainly falls into two categories: traditional static planning methods and emerging learning-based dynamic methods. Traditional methods typically assume a known and fixed network traffic model, making them difficult to adapt to complex and ever-changing real-world conditions, and they also incur extremely high computational complexity. In recent years, deep reinforcement learning has been introduced into this field due to its powerful online decision-making capabilities. However, existing DRL-based TSN scheduling schemes still face two major challenges in practical applications:

[0004] Most existing DRL scheduling algorithms are based on an ideal assumption: the scheduler can obtain global, accurate, and complete network state information in real time. However, in real-world operating environments, due to factors such as electromagnetic interference, momentary link jitter, and occasional sensor or node failures, the data observed by the scheduler is often an "irregular time series" containing noise and missing values. Under these "non-ideal observation" conditions, traditional DRL models cannot accurately perceive the environmental state, leading to a sharp decline in decision-making performance or even failure, seriously threatening network reliability.

[0005] Traffic flow scheduling networks are typical multi-service converged networks, carrying traffic of different critical levels. For example, in autonomous driving scenarios, vehicle control signals belong to Time-Trigger (TT) flows, which must be guaranteed to arrive before the deadline; while entertainment data belongs to Best-Effect (BE) flows, which are not sensitive to latency but pursue high throughput. Existing DRL methods typically use a single scalar reward function for optimization. This strategy, lacking hierarchical differentiation, easily leads to the model sacrificing the deterministic guarantee of critical flows in order to reduce the overall average latency, which is unacceptable in real-world scenarios. Summary of the Invention

[0006] To overcome the shortcomings of existing technologies and compensate for their poor adaptability in non-ideal observation environments and the inability of a single optimization objective to meet various differentiated business needs, this invention proposes a TSN traffic scheduling algorithm and system based on causal data completion. This aims to solve the scheduling failure problem caused by data loss due to network jitter or node failure. The technical solution adopted by this invention is as follows: a time-sensitive network traffic scheduling method based on causal data completion is used to collect real-time traffic of the current time-sensitive TSN network and construct an irregular observation matrix; missing data in the observation matrix is ​​completed using a state reconstruction module based on neural Granger causality; robust spatiotemporal features with noise resistance are extracted from the completed observation matrix through a variational graph attention network; these robust spatiotemporal features are input into a reinforcement learning decision network based on hierarchical optimization, and the best-effort BE flow transmission efficiency is optimized under the hard real-time constraint of priority critical time-triggered TT flow, ultimately generating a gated control list for execution.

[0007] The specific steps are as follows:

[0008] Step 1: Obtain real-time traffic status data of the time-sensitive TSN network and construct an irregular observation matrix containing time series data, wherein the observation matrix contains missing data due to network jitter or node transient failures.

[0009] Step 2: Construct a state reconstruction module based on neural Granger causality, and use an alternating optimization strategy of latent data prediction and causal graph fitting to complete the irregular observation matrix and output the complete network state matrix.

[0010] Step 3: Construct a variational feature extraction module based on the variational graph attention network V-GAT, map the complete network state matrix output in Step 2 to the latent feature distribution, and sample the state feature vector from it;

[0011] Step 4: Construct a hierarchical optimized deep reinforcement learning decision model, calculate the reward function based on the state feature vector obtained in Step 3 (i.e., traffic type distinction), and output the scheduling action;

[0012] Step 5: Generate a gating control list (GCL) based on the scheduling action and send it to the time-sensitive TSN network switch to perform traffic scheduling.

[0013] In step 1, the TSN network topology is abstracted into a directed graph. ,in This represents the set of terminal systems and switch nodes. Represents the set of physical links; defines the time. The network traffic state matrix is ,in For the number of nodes, The feature dimension includes traffic size, period, remaining transmission time, and queue length; a binary mask matrix is ​​defined. When node At any moment Data loss due to link jitter or failure =0, otherwise =1; Construct irregular time series data as actually observed. Its mathematical expression is:

[0014]

[0015] Among them, symbols This represents the Hadamard product, which is the element-wise multiplication of matrices, with unobserved data positions set to 0.

[0016] In step 2, the state reconstruction module fills in the gaps by learning the causal relationships in the historical node traffic. To address missing values ​​in the graph, a delayed supervised graph neural network is first designed to predict node states, with historical observation data as input. and the current causal probability diagram The prediction formula is:

[0017]

[0018] Among them Maximum time lag;

[0019] The module then learns from node j to node j through cause-effect graph fitting. The probability of causal influence is determined by updating the network parameters and causal graph structure by minimizing the prediction error:

[0020]

[0021] in To constrain sparsity, ensure that the learned causal graph is sparse and efficient;

[0022] Finally, the trained model is used to fill in the gaps. For missing values ​​at each location, output the complete network state matrix. :

[0023]

[0024] In step 3, the Variational Graph Attention Network (V-GAT) is introduced for feature extraction. By introducing the concept of variational autoencoders, the node features are modeled as a probability distribution. The specific steps are as follows:

[0025] First, graph attention is used to calculate the degree of association between nodes:

[0026]

[0027] in and This represents the original state vector between nodes. Represents a linear transformation matrix. This represents a vector concatenation operation. Represents the weight vector of the attention mechanism;

[0028] V-GAT uses neighbor information to infer the mean and standard deviation of the current node's feature distribution, specifically using two independent weight matrices. and :

[0029]

[0030]

[0031] in The mean vector representing the distribution of latent features of nodes. The variance vector representing the distribution of latent features of nodes. Represents a node The set of all neighboring nodes;

[0032] To obtain the final feature vector input to the reinforcement learning network while preserving the differentiability of the gradient, a reparameterization technique is employed:

[0033]

[0034] in This is the final robust latent feature vector. From normal distribution Random noise in the mid-sample;

[0035] Even if the input data has a slight perturbation It can also maintain the stability of the distribution, thus making subsequent scheduling strategies more robust;

[0036] In step 4, hierarchical optimization is used to design the target and reward functions based on the priority of transmission traffic;

[0037] A hybrid traffic scheduling model is adopted, prioritizing the scheduling of important TT flows, with the aim of minimizing the following hierarchical objective function:

[0038]

[0039] in, For a set of latency-sensitive critical traffic, For a collection of ordinary traffic that does its best, This indicates the penalty for TT stream timeout. This represents the total queuing time for the BE flow. and For the weighting coefficients, it is required that... ;

[0040] In reinforcement learning networks, the reward an agent receives at each step depends on the type of traffic currently being scheduled. The reward function is calculated dynamically, and the formula is as follows:

[0041]

[0042] in Indicates the current time The data stream that is selected for action scheduling. It is a huge complex number representing the enormous penalty for timeout. Indicates scheduling gain:

[0043]

[0044] In step 5, the extracted features are... The input is fed into the reinforcement learning Actor decision network, which makes specific scheduling decisions based on the output actions. This series of actions is then mapped to a gating control list defined by the IEEE 802.1Qbv standard and sent to the registers of each TSN switch through the control channel to complete the traffic scheduling execution of the physical link.

[0045] The TSN traffic scheduling system based on causal data completion includes a sensor group, a TSN switch, a central computing platform, and a display terminal. First, the sensor group generates periodic or bursty traffic requests. The TSN switch is responsible for transmitting this traffic over the physical link. The central computing platform, as the core execution unit of the system, obtains the entire network topology status and traffic information through its built-in bus interface and sends the calculated gating control list to the TSN switch for execution. The central computing platform includes the following modules:

[0046] Data sensing module: used to construct the network state matrix, mask matrix, and irregular time series observation set to be scheduled;

[0047] State Reconstruction Module: This module has a built-in data completion algorithm unit. Based on the neural Granger causality state reconstruction module, this unit uses an alternating optimization strategy of latent data prediction and causal graph fitting to complete the irregular observation matrix. Based on the observation data to be detected and the mask, it repairs the irregular time series data and outputs the complete network state.

[0048] Variational feature extraction module: Based on the reconstructed complete state, extract important features and output the optimal traffic scheduling action;

[0049] Execution module: Converts the scheduling policy into a TSN protocol-compatible gated list (GCL) and issues it for execution.

[0050] The features and beneficial effects of this invention are:

[0051] (1) This invention proposes a TSN traffic scheduling algorithm based on causal data completion. By constructing a state reconstruction module based on neural Granger causality, the missing data is accurately filled in without relying on prior knowledge. This breaks through the limitation of traditional scheduling algorithms that rely too much on global ideal observation and ensures the availability of the scheduler in harsh communication environments.

[0052] (2) A deep reinforcement learning decision-making system based on hierarchical optimization was constructed, which solved the problem that a single optimization objective could not meet the differentiated needs of multiple services. By designing a hierarchical reward function that includes hard constraints (for TT streams) and soft optimization (for BE streams), the throughput and transmission efficiency of ordinary entertainment traffic were maximized under the premise of absolutely prioritizing the hard real-time performance and deterministic transmission of key control signals. Attached Figure Description

[0053] Figure 1 A schematic diagram of a TSN traffic scheduling algorithm based on causal data completion.

[0054] Figure 2 : Schematic diagram of the feature extraction module of variational graph attention network.

[0055] Figure 3 : Interaction diagram of deep reinforcement learning decision module based on hierarchical optimization. Detailed Implementation

[0056] This invention is mainly applied to high-reliability traffic scheduling scenarios in communication networks.

[0057] The main contents of this invention are as follows: real-time network traffic is collected and an irregular observation matrix is ​​constructed; missing data is completed using a state reconstruction module based on neural Granger causality; robust spatiotemporal features with noise resistance are extracted through a variational graph attention network; these features are input into a reinforcement learning decision model based on hierarchical optimization, and the transmission efficiency of the BE stream is optimized under the hard real-time constraint of priority TT stream, and finally a gated control list is generated and executed.

[0058] The technical solution proposed in this invention is as follows: a time-sensitive network traffic scheduling method based on causal data completion is used to collect the real-time traffic of the current time-sensitive network (TSN) and construct an irregular observation matrix; the missing data in the observation matrix is ​​completed using a state reconstruction module based on neural Granger causality; robust spatiotemporal features with noise resistance are extracted from the completed observation matrix through a variational graph attention network; the robust spatiotemporal features with noise resistance are input into a reinforcement learning decision network based on hierarchical optimization, and the best-effort BE flow transmission efficiency is optimized under the hard real-time constraint of priority critical time triggering TT flow, and finally a gated control list is generated and executed.

[0059] The specific steps are as follows:

[0060] Step 1: Obtain real-time traffic status data of the time-sensitive TSN network and construct an irregular observation matrix containing time series data, wherein the observation matrix contains missing data due to network jitter or node transient failures.

[0061] Step 2: Construct a state reconstruction module based on neural Granger causality, and use an alternating optimization strategy of latent data prediction and causal graph fitting to complete the irregular observation matrix and output the complete network state matrix.

[0062] Step 3: Construct a variational feature extraction module based on the variational graph attention network V-GAT, map the complete network state matrix output in Step 2 to the latent feature distribution, and sample the state feature vector from it;

[0063] Step 4: Construct a hierarchical optimized deep reinforcement learning decision model, calculate the reward function based on the state feature vector obtained in Step 3 (i.e., traffic type distinction), and output the scheduling action;

[0064] Step 5: Generate a gating control list (GCL) based on the scheduling action and send it to the time-sensitive TSN network switch to perform traffic scheduling.

[0065] In step 1, the TSN network topology is abstracted into a directed graph. ,in This represents the set of terminal systems and switch nodes. Represents the set of physical links; defines the time. The network traffic state matrix is ,in For the number of nodes, The feature dimension includes traffic size, period, remaining transmission time, and queue length; a binary mask matrix is ​​defined. When node At any moment Data loss due to link jitter or failure =0, otherwise =1; Construct irregular time series data as actually observed. Its mathematical expression is:

[0066]

[0067] Among them, symbols This represents the Hadamard product, which is the element-wise multiplication of matrices, with unobserved data positions set to 0.

[0068] In step 2, the state reconstruction module fills in the gaps by learning the causal relationships in the historical node traffic. To address missing values ​​in the graph, a delayed supervised graph neural network is first designed to predict node states, with historical observation data as input. and the current causal probability diagram The prediction formula is:

[0069]

[0070] Among them Maximum time lag;

[0071] The module then learns from node j to node j through cause-effect graph fitting. The probability of causal influence is determined by updating the network parameters and causal graph structure by minimizing the prediction error:

[0072]

[0073] in To constrain sparsity, ensure that the learned causal graph is sparse and efficient;

[0074] Finally, the trained model is used to fill in the gaps. For missing values ​​at each location, output the complete network state matrix. :

[0075]

[0076] In step 3, the Variational Graph Attention Network (V-GAT) is introduced for feature extraction. By introducing the concept of variational autoencoders, the node features are modeled as a probability distribution. The specific steps are as follows:

[0077] First, graph attention is used to calculate the degree of association between nodes:

[0078]

[0079] in and This represents the original state vector between nodes. Represents a linear transformation matrix. This represents a vector concatenation operation. Represents the weight vector of the attention mechanism;

[0080] V-GAT uses neighbor information to infer the mean and standard deviation of the current node's feature distribution, specifically using two independent weight matrices. and :

[0081]

[0082]

[0083] in The mean vector representing the distribution of latent features of nodes. The variance vector representing the distribution of latent features of nodes. Represents a node The set of all neighboring nodes;

[0084] To obtain the final feature vector input to the reinforcement learning network while preserving the differentiability of the gradient, a reparameterization technique is employed:

[0085]

[0086] in This is the final robust latent feature vector. From normal distribution Random noise in the mid-sample;

[0087] Even if the input data has a slight perturbation It can also maintain the stability of the distribution, thus making subsequent scheduling strategies more robust;

[0088] In step 4, hierarchical optimization is used to design the target and reward functions based on the priority of transmission traffic;

[0089] A hybrid traffic scheduling model is adopted, prioritizing the scheduling of important TT flows, with the aim of minimizing the following hierarchical objective function:

[0090]

[0091] in, For a set of latency-sensitive critical traffic, For a collection of ordinary traffic that does its best, This indicates the penalty for TT stream timeout. This represents the total queuing time for the BE flow. and For the weighting coefficients, it is required that... ;

[0092] In reinforcement learning networks, the reward an agent receives at each step depends on the type of traffic currently being scheduled. The reward function is calculated dynamically, and the formula is as follows:

[0093]

[0094] in Indicates the current time The data stream that is selected for action scheduling. It is a huge complex number representing the enormous penalty for timeout. Indicates scheduling gain:

[0095]

[0096] In step 5, the extracted features are... The input is fed into the reinforcement learning Actor decision network, which makes specific scheduling decisions based on the output actions. This series of actions is then mapped to a gating control list defined by the IEEE 802.1Qbv standard and sent to the registers of each TSN switch through the control channel to complete the traffic scheduling execution of the physical link.

[0097] The TSN traffic scheduling system based on causal data completion includes a sensor group, a TSN switch, a central computing platform, and a display terminal. First, the sensor group generates periodic or bursty traffic requests. The TSN switch is responsible for transmitting this traffic over the physical link. The central computing platform, as the core execution unit of the system, obtains the entire network topology status and traffic information through its built-in bus interface and sends the calculated gating control list to the TSN switch for execution. The central computing platform includes the following modules:

[0098] Data sensing module: used to construct the network state matrix, mask matrix, and irregular time series observation set to be scheduled;

[0099] State Reconstruction Module: This module has a built-in data completion algorithm unit. Based on the neural Granger causality state reconstruction module, this unit uses an alternating optimization strategy of latent data prediction and causal graph fitting to complete the irregular observation matrix. Based on the observation data to be detected and the mask, it repairs the irregular time series data and outputs the complete network state.

[0100] Variational feature extraction module: Based on the reconstructed complete state, extract important features and output the optimal traffic scheduling action;

[0101] Execution module: Converts the scheduling policy into a TSN protocol-compatible gated list (GCL) and issues it for execution.

[0102] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0103] The purpose of this invention is to propose a TSN traffic scheduling algorithm design based on causal data completion. First, network traffic and topology status are collected in real time through gateway and switch interfaces to construct an irregular time-series observation matrix containing mask identifiers. Then, a state reconstruction module performs self-supervised repair on the incomplete observation data, outputting a complete network state matrix. A variational graph attention network is used to extract deep features from the reconstructed state, which is then input into a hierarchical optimization-based reinforcement learning agent. Finally, a standard-compliant gating control list is generated based on the agent's output actions and distributed to the registers of each TSN switch. This method can achieve highly reliable traffic scheduling control under non-ideal observation conditions.

[0104] The specific implementation steps are as follows:

[0105] Step 1: Obtain real-time network traffic status data and construct an irregular observation matrix containing time series data, wherein the observation matrix includes missing data due to network jitter or transient node failures.

[0106] Step 2: Construct a state reconstruction module based on neural Granger causality, and use an alternating optimization strategy of latent data prediction and causal graph fitting to complete the irregular observation matrix and output the complete network state matrix.

[0107] Step 3: Construct a variational feature extraction module based on Variational Graph Attention Network (V-GAT), map the complete network state matrix output in Step 2 to the latent feature distribution, and sample the state feature vector from it;

[0108] Step 4: Construct a hierarchical optimization deep reinforcement learning decision model, calculate the reward function according to the traffic type, and output the scheduling action;

[0109] Step 5: Generate a gated control list (GCL) based on the scheduling action and send it to the TSN switch to perform traffic scheduling.

[0110] The data perception module abstracts the perceived TSN network topology into a directed graph. ,in This represents the set of terminal systems and switch nodes. Represents the set of physical links; defines the time. The network traffic state matrix is ,in For the number of nodes, The feature dimension includes traffic size, period, remaining transmission time, and queue length; a binary mask matrix is ​​defined. When node At any moment Data loss due to link jitter or failure =0, otherwise =1; Construct irregular time series data as actually observed. Its mathematical expression is:

[0111]

[0112] Among them, symbols This represents the Hadamard product, which is the element-wise multiplication of matrices, with unobserved data positions set to 0.

[0113] The state reconstruction module is a data repair architecture designed for non-ideal network observation environments. It leverages a deep learning algorithm based on neural Granger causality to uncover the latent spatiotemporal dependencies in inter-node traffic, overcoming the limitations of traditional scheduling systems that rely on complete observation data. Simultaneously, it employs an alternating optimization strategy of latent data prediction and causal structure learning to achieve high-precision state completion even when data loss occurs due to network jitter or node failures, providing a complete and robust global state input for downstream scheduling decisions.

[0114] By learning the causal relationships in historical node traffic, fill in the gaps. To address missing values ​​in the graph, a delayed supervised graph neural network is first designed to predict node states, with historical observation data as input. and the current causal probability diagram The prediction formula is:

[0115]

[0116] Among them Maximum time lag;

[0117] The module learns the path from node j to node j by fitting a causal graph. The probability of causal influence is determined by updating the network parameters and causal graph structure by minimizing the prediction error:

[0118]

[0119] in To constrain sparsity, ensure that the learned causal graph is sparse and efficient;

[0120] Finally, the trained model is used to fill in the gaps. For missing values ​​at each location, output the complete network state matrix. ;

[0121]

[0122] The variational feature extraction module extracts features by introducing a variational graph attention network. First, graph attention is used to calculate the degree of association between nodes:

[0123]

[0124] in and This represents the original state vector between nodes. Represents a linear transformation matrix. This represents a vector concatenation operation. Represents the weight vector of the attention mechanism;

[0125] V-GAT uses neighbor information to infer the mean and standard deviation of the current node's feature distribution. We use two independent weight matrices. and :

[0126]

[0127]

[0128] in The mean vector representing the distribution of latent features of nodes. The variance vector representing the distribution of latent features of nodes. Represents a node The set of all neighboring nodes;

[0129] To obtain the final feature vector input to the reinforcement learning network while preserving the differentiability of the gradient, a reparameterization technique is employed:

[0130]

[0131] in This is the final robust latent feature vector. From normal distribution Random noise from mid-sampled data; after feature extraction, the system designs a hierarchical optimization reward function based on the priority of transmission traffic for traffic scheduling. The objective and reward function are as follows:

[0132]

[0133]

[0134] in, For a set of latency-sensitive critical traffic, For a collection of ordinary traffic that does its best, This indicates the penalty for TT stream timeout. This represents the total queuing time for the BE flow. and For the weighting coefficients, it is required that... , Indicates the current time The data stream that is selected for action scheduling. It is a huge complex number representing the enormous penalty for timeout. Indicates scheduling gain: .

[0135] The execution module will feature The Actor network, which is input into reinforcement learning, makes specific scheduling decisions based on the output actions. It maps these actions to a gating control list defined by the IEEE 802.1Qbv standard and sends them to the registers of each TSN switch through the control channel to complete the traffic scheduling execution of the physical link.

[0136] The complete training process of this invention is as follows:

[0137] a. Network parameter initialization: Set the number of training epochs to 1000, and the maximum time step per epoch. Set one supercycle of TSN and initialize the prediction network parameters. Variational attention network weight matrix Attention vector Tiered reward weight , TT stream timeout penalty coefficient ;

[0138] b. Data Loading: Read the network topology file and construct a directed graph. Define the node set and physical link set, and introduce a random masking mechanism into the training data: using probabilities in the range of 0.1-0.3. Randomly select some nodes' observation states Set to 0 to generate an irregular observation matrix. This is used to train the model's noise resistance and completion capabilities. State features such as traffic volume, remaining transmission time, and queue length are normalized to the [0,1] interval.

[0139] c. State reconstruction training: (The masked state is then reconstructed.) Input the prediction network, use historical node data to predict the current missing value, and alternately optimize the prediction network and the causal graph structure to minimize the prediction error. and sparsity penalty ;

[0140] d. Feature extraction and sampling: The completed state... Input the V-GAT (Variational GAT) module, which uses a graph attention mechanism to calculate the attention coefficients between nodes and infer the mean of node features. and logarithmic variance Normal noise is introduced for reparameterization sampling to generate robust feature components. ;

[0141] e. Hierarchical reinforcement learning interaction: ① Integrating feature components ① Input the reinforcement learning Actor network and output the scheduling action; ② The environment executes the action and calculates the hierarchical reward. If the current scheduling is a TT flow and... This triggers a hard constraint penalty. If the queue time does not expire or the process is a BE (Best Before End) flow, a reward will be given based on the reduction in queue time. ③ Collect action rewards and store them in the experience replay pool;

[0142] f. Model Validation and Saving: Every 10 rounds of training, the model is tested on a preset validation set. The default rate of the TT stream and the average latency of the BE stream are calculated. When the default rate of the TT stream is 0 and the total reward exceeds the historical best value, the model parameters of all modules are saved.

[0143] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A time-sensitive network traffic scheduling method based on causal data completion, characterized in that, Real-time traffic of the current time-sensitive TSN network is collected and an irregular observation matrix is ​​constructed. Missing data in the observation matrix is ​​completed using a state reconstruction module based on neural Granger causality. Robust spatiotemporal features with noise resistance are extracted from the completed observation matrix through a variational graph attention network. The robust spatiotemporal features with noise resistance are input into a reinforcement learning decision network based on hierarchical optimization. The best-effort BE stream transmission efficiency is optimized under the hard real-time constraint of priority key time triggering TT stream, and finally a gating control list is generated and executed.

2. The time-sensitive network traffic scheduling method based on causal data completion as described in claim 1, characterized in that, The specific steps are as follows: Step 1: Obtain real-time traffic status data of the time-sensitive TSN network and construct an irregular observation matrix containing time series data, wherein the observation matrix contains missing data due to network jitter or node transient failures. Step 2: Construct a state reconstruction module based on neural Granger causality, and use an alternating optimization strategy of latent data prediction and causal graph fitting to complete the irregular observation matrix and output the complete network state matrix. Step 3: Construct a variational feature extraction module based on the variational graph attention network V-GAT, map the complete network state matrix output in Step 2 to the latent feature distribution, and sample the state feature vector from it; Step 4: Construct a hierarchical optimized deep reinforcement learning decision model, calculate the reward function based on the state feature vector obtained in Step 3 (i.e., traffic type distinction), and output the scheduling action; Step 5: Generate a gating control list (GCL) based on the scheduling action and send it to the time-sensitive TSN network switch to perform traffic scheduling.

3. The time-sensitive network traffic scheduling method based on causal data completion as described in claim 1, characterized in that, In step 1, the TSN network topology is abstracted into a directed graph. ,in This represents the set of terminal systems and switch nodes. Represents the set of physical links; defines the time. The network traffic state matrix is ,in For the number of nodes, The feature dimension includes traffic size, period, remaining transmission time, and queue length; a binary mask matrix is ​​defined. When node At any moment Data loss due to link jitter or failure =0, otherwise =1; Constructing irregular time series data observed in practice Its mathematical expression is: Among them, symbols This represents the Hadamard product, which is the element-wise multiplication of matrices, with unobserved data positions set to 0. In step 2, the state reconstruction module fills in the gaps by learning the causal relationships in the historical node traffic. To address missing values ​​in the graph, a delayed supervised graph neural network is first designed to predict node states, with historical observation data as input. and the current causal probability diagram The prediction formula is: Among them Maximum time lag; The module then learns from node j to node j through cause-effect graph fitting. The probability of causal influence is determined by updating the network parameters and causal graph structure by minimizing the prediction error: in To constrain sparsity, ensure that the learned causal graph is sparse and efficient; Finally, the trained model is used to fill in the gaps. For missing values ​​at each location, output the complete network state matrix. : In step 3, the Variational Graph Attention Network (V-GAT) is introduced for feature extraction. By introducing the concept of variational autoencoders, the node features are modeled as a probability distribution. The specific steps are as follows: First, graph attention is used to calculate the degree of association between nodes: in and This represents the original state vector between nodes. Represents a linear transformation matrix. This represents a vector concatenation operation. Represents the weight vector of the attention mechanism; V-GAT uses neighbor information to infer the mean and standard deviation of the current node's feature distribution, specifically using two independent weight matrices. and : in The mean vector representing the distribution of latent features of nodes. The variance vector representing the distribution of latent features of nodes. Represents a node The set of all neighboring nodes; To obtain the final feature vector input to the reinforcement learning network while preserving the differentiability of the gradient, a reparameterization technique is employed: in This is the final robust latent feature vector. From normal distribution Random noise in the mid-sample; Even if the input data has a slight perturbation It can also maintain the stability of the distribution, thus making subsequent scheduling strategies more robust; In step 4, hierarchical optimization is used to design the target and reward functions based on the priority of transmission traffic; A hybrid traffic scheduling model is adopted, prioritizing the scheduling of important TT flows, with the aim of minimizing the following hierarchical objective function: in, For a set of latency-sensitive critical traffic, For a collection of ordinary traffic that does its best, This indicates the penalty for TT stream timeout. This represents the total queuing time for the BE flow. and For the weighting coefficients, it is required that... ; In reinforcement learning networks, the reward an agent receives at each step depends on the type of traffic currently being scheduled. The reward function is calculated dynamically, and the formula is as follows: in Indicates the current time The data stream that is selected for action scheduling. It is a huge complex number representing the enormous penalty for timeout. Indicates scheduling gain: In step 5, the extracted features are... The input is fed into the reinforcement learning Actor decision network, which makes specific scheduling decisions based on the output actions. This series of actions is then mapped to a gating control list defined by the IEEE 802.1Qbv standard and sent to the registers of each TSN switch through the control channel to complete the traffic scheduling execution of the physical link.

4. A TSN traffic scheduling system based on causal data completion, characterized in that, The system includes sensor arrays, TSN switches, a central computing platform, and display terminals. First, the sensor arrays generate periodic or bursty traffic requests. The TSN switches are responsible for transmitting this traffic over the physical link. The central computing platform, as the core execution unit of the system, obtains the overall network topology status and traffic information through its built-in bus interface and sends the calculated gating control list to the TSN switches for execution. The central computing platform includes the following modules: Data sensing module: used to construct the network state matrix, mask matrix, and irregular time series observation set to be scheduled; State Reconstruction Module: This module has a built-in data completion algorithm unit. Based on the neural Granger causality state reconstruction module, this unit uses an alternating optimization strategy of latent data prediction and causal graph fitting to complete the irregular observation matrix. Based on the observation data to be detected and the mask, it repairs the irregular time series data and outputs the complete network state. Variational feature extraction module: Based on the reconstructed complete state, extract important features and output the optimal traffic scheduling action; Execution module: Converts the scheduling policy into a TSN protocol-compatible gated list (GCL) and issues it for execution.

Images