Spatial characteristic enhanced Informer network data transmission deterministic time delay guarantee method

Abstract

The invention discloses a network data transmission deterministic time delay guarantee method of a spatial characteristic enhanced Informer model, which comprises the following steps of: 1, predicting a time sequence of key service traffic based on the Informer model, and extracting a long-period time characteristic and a future traffic trend through a Prob sparse self-attention mechanism; 2, a graph neural network is introduced into the model to perform modeling on an underground switch topological structure, the defect that Informer only depends on time characteristics and neglects network topological information is improved, and a spatial state matrix is generated and is combined with a future flow trend to realize enhancement of spatial characteristics; 3, an asynchronous scheduling mechanism is designed, prediction calculation and short-period scheduling execution are decoupled, and real-time performance is guaranteed; and 4, generating an optimal step traffic shaper parameter by using the spatio-temporal characteristics to smooth burst traffic, allocate bandwidth and reduce delay jitter, and meanwhile, combining burst detection and redundancy scheduling protection to guarantee deterministic delay. The method is suitable for industrial networks with high real-time requirements such as coal mines and rail transit.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial network communication and data scheduling, and particularly relates to a network data transmission deterministic delay guarantee method of a space characteristic enhanced Informer. BACKGROUND

[0002] Time-sensitive network (TSN) is a real-time communication technology system developed on the basis of standard Ethernet. It introduces mechanisms such as time synchronization, traffic scheduling, and priority control based on IEEE 802.1 series standards, aiming to solve the needs of low latency, predictability, and high reliability. TSN can achieve ordered transmission of data with microsecond-level delay, and through priority queue and shaping mechanisms, it guarantees the stability of critical flows in a multi-service mixed network environment, and thus has been widely applied in the fields of industrial automation, smart grid, rail transit, aerospace, etc.

[0003] Among the various traffic control and scheduling mechanisms provided by TSN, asynchronous traffic shaping (ATS) is a key technology that does not require global clock synchronization. Its core idea is to calculate the frame eligibility time of high-priority scheduled traffic and update the state machine variables at the ingress port of the switch, thereby smoothing burst data and achieving end-to-end delay control and queue jitter suppression. Compared with synchronous scheduling mechanisms, ATS has low dependence on time synchronization accuracy and high deployment flexibility, and is more suitable for industrial networks with complex topology and widely distributed nodes. Common ATS algorithms include LRQ, TBE, ATS native algorithm, and Paternoster, which adjust traffic rate and priority allocation by configuring parameters such as token bucket capacity, filling rate, and gating list, thereby ensuring low latency of critical flows while improving link utilization.

[0004] However, the existing ATS parameter configuration method still has the following deficiencies: 1. Insufficient feature utilization and lack of spatio-temporal joint modeling capability: Real network traffic contains both spatial features (topology structure, link bandwidth, utilization rate, end-to-end delay, etc.) and temporal features (periodic fluctuations, burst patterns, etc.), but traditional methods are mostly single-dimensional modeling, which fails to effectively integrate both types of features, resulting in a lack of globality and foresight in parameter optimization.

[0005] 2. Poor dynamic adaptability: Industrial field network environments fluctuate frequently, with both periodic changes in load and sudden events such as link failures, device switching, network attacks, etc. Fixed rules or ATS configurations relying on short-term historical data are difficult to respond to these changes in a timely manner.

[0006] 3. Real-time and reliability of prediction and scheduling are difficult to balance: high-precision prediction and complex optimization algorithm have large calculation overhead, which is difficult to complete within a millisecond scheduling period, resulting in that the prediction result cannot be applied in real time; and when the prediction is inaccurate or fails, the existing method lacks the ability to quickly detect and switch to a safe scheduling strategy, which easily causes delay of critical service flow to exceed the standard and the packet loss rate to rise.

[0007] Therefore, based on the above status quo, there is an urgent need for a network data transmission deterministic latency guarantee method for spatial characteristic enhanced Informer to change this status quo. SUMMARY

[0008] The purpose of the present application is to solve the shortcomings in the prior art and propose a network data transmission deterministic latency guarantee method for spatial characteristic enhanced Informer. Its advantages lie in balancing scheduling accuracy, real-time performance and system robustness, and it is suitable for industrial network environments such as coal mines and rail transit that have strict delay requirements.

[0009] In order to achieve the above purpose, the present application adopts the following technical solutions: A network data transmission deterministic latency guarantee method for spatial characteristic enhanced Informer, comprising the following steps: Step 1: Based on the Informer model, the traffic of multiple types of key business traffic GOOSE, SV, MMS, etc. is predicted; a multivariate time series is constructed by historical traffic , and long-period time features are extracted using ProbSparse self-attention mechanism to predict the traffic trend in multiple future scheduling periods ; Step 2: Based on the graph neural network, a spatial characteristic model of the TSN network is constructed, taking switches as a node set , links as an edge set , node features including link bandwidth , utilization rate , latency , bit error rate and node degree , the attention weight of the node and its neighbor nodes is calculated through the graph attention network to generate a spatial state matrix reflecting the real-time transmission capability of the network ; then, the spatial state matrix is used to enhance the spatial characteristics of the model in step 1, and the future traffic sequence is fused with the spatial state matrix to obtain a spatio-temporal feature vector for subsequent input ; Step 3: Based on the obtained spatio-temporal feature vector , the off-line trained network is used to generate optimal asynchronous traffic shaper token bucket capacity , filling rate and gating list with the optimization goal of minimizing key service end-to-end average delay, jitter and packet loss rate; after training, a state-action mapping table is generated, which is used in the subsequent online stage to obtain the optimal parameters of the current scheduling period in a millisecond-level table lookup manner; Step 4: Set the asynchronous scheduling mechanism to decouple the prediction calculation and the short-period scheduling execution, and calculate the optimal asynchronous traffic shaper parameter sequence covering multiple short periods in advance within the prediction window for smoothing burst traffic, allocating bandwidth and reducing delay jitter; in the online execution stage, the corresponding parameters are directly called by table lookup in each short period; Step 5: Set a burst detection algorithm during operation, based on sliding window statistical calculation to monitor real-time traffic surge, link anomaly or prediction mismatch, and switch to pre-set conservative traffic shaping parameters when the trigger condition is met to reserve bandwidth and priority for key service flows, and switch back to the optimal parameter configuration after the prediction module recovers to normal.

[0010] The application further sets that the step 1 is performed as follows: Firstly, the input data of the model is the historical time series data of multiple key service traffic types GOOSE, SV and MMS, i.e. the key service flow set ; these traffic data represent the transmission situation of different service types in the network, containing the time sequence variation characteristics; Input data: time series data , including multi-dimensional time series data of historical traffic; ; Among them, is the observation data of GOOSE, SV and MMS traffic at a fixed time step ; each service flow is sampled at a sampling period to obtain the traffic sequence of the past time points: ; Among them, the service traffic type ; the multi-service flow is spliced into an input tensor: , M=| |; Among them, is the total number of service traffic types; Received via Informer model The data is encoded to enable traffic prediction for multiple future scheduling cycles of time-sensitive networks in coal mines. The Informer model employs a multi-head probabilistic sparse self-attention mechanism to uncover potential relationships between vectors and provide prediction results for long-term series in one go. The present invention is further configured such that the Informer model mainly consists of an encoder and a decoder; the data first enters the embedding layer of the encoder, which contains three encoding modules: an embedding encoding module, a position encoding module, and a timestamp encoding module. Each module encodes the input data, and the encoded data are concatenated to obtain the encoded feature matrix. ; Next, the encoded feature matrix Processed into 3 matrices of the same size, namely the query matrix. Key matrix Value matrix They are respectively Sequence length, For sequence feature dimensions: ; In the formula, They are respectively The weight matrix; Through the key matrix Random sampling, obtaining the sampled At this point, query matrix The i-th row The corresponding sparsity metric is: ; in, It is a sparsity metric used to determine the importance of the query vector in attention calculation; It is a query matrix The i-th row, for The OK, For sequence feature dimensions, It is the sequence length of the sampled key matrix. It takes the maximum value of all scores in the sample key set; Select the corresponding sparsity metric with a higher value To form a new query matrix Calculate probabilistic sparse self-attention: ; in, It is the dimension of the query vector. The transpose of the key matrix determines the scale of attention computation, and the attention weights are normalized through the Softmax operation. To enhance the model's ability to capture long-term dependencies in long-sequence data while effectively reducing computational complexity, multi-head attention mechanisms concatenate the outputs of multiple different attention heads, integrating multi-dimensional traffic information: ; in, It is the overall output of the multi-head attention mechanism. It is the output of multiple heads, representing the focus on different features; For splicing operations, It is a linear transformation matrix that generates the final prediction result; To ensure computational efficiency, interlayer distillation is used to extract core features from each layer, and these features are then used as a basis to build a more efficient and focused self-attention representation in the next layer; the encoder distillation layer receives the first... The output of the layer after multi-head probabilistic sparse self-attention operation; this layer contains one-dimensional convolution and max pooling operations to achieve dimensionality compression; from the th Feature extraction from layer 1 is passed to layer 2. The layer operation process is as follows: ; in, For the first time after distillation layer The feature matrix at time step; It is the max pooling function; for Activation function; It is a one-dimensional convolutional filter; For the first layer Time-time feature matrix; This represents a multi-head probability sparse self-attention operation; The present invention is further configured such that the Informer model decoder consists of two multi-head self-attention layers; by introducing a generative decoder and a masking mechanism, the traffic prediction result depends only on the data before the sequence to be predicted; under this mechanism, the data of the sequence to be predicted is masked to ensure that the model only focuses on the known data part; the input of the decoder is the concatenation of the temporal feature matrix from the encoder and the sequence to be predicted; The encoder and decoder interact via a cross-attention mechanism. The decoder utilizes the global feature information extracted by the encoder, combined with the contextual dependencies of the target sequence, to construct a feature matrix. This feature matrix is ​​then input into the fully connected layer to output the predicted traffic value. The output predicted sequence is: ; in, This refers to the total number of business traffic types. The desired time step is predicted, so asynchronous prediction is used to ensure real-time performance. To ensure the model can handle different types of traffic simultaneously, the loss function needs to consider the weight of each type of traffic and calculate the prediction error using mean squared error; specifically, the loss function... By calculating each type of business traffic ( The mean squared error between the predicted and actual values ​​is calculated, then a weighted average is taken, and finally the errors for all traffic types are summed; the loss function is calculated as follows: ; in, It is the traffic of each type of business. The weighting coefficients are typically assigned values ​​based on the importance of different traffic types. ; It predicts the time step. It is the model's prediction of the first Time of the first Predicted values ​​for traffic types, It is the actual first model Time of the first The actual value of traffic.

[0011] The present invention is further configured such that step 2 is performed according to the following steps: First, a graph neural network is used to model the switch topology in a Time-Sensitive Network (TSN); the switches in the network are treated as nodes. A link is considered as an edge between nodes. Network topology diagram Define the characteristics of each node, including link bandwidth. Utilization rate Delay Bit error rate and node degree Parameters; for each switch node Its eigenvectors include: ; in, This indicates the link bandwidth, reflecting the link's transmission capacity; Indicates link utilization; This represents the link delay between node i and its neighbor node j; Bit error rate (BER) represents the probability that a data packet will encounter an error during transmission; while Represents a node The degree of these nodes; these node features are used as input to the graph neural network for further processing; Next, network features are extracted using the Graph Attention Network (GAT) mechanism in graph neural networks. In a graph attention network, the relationships between nodes, i.e., the degree of influence each node has on its neighboring nodes, are calculated through the attention mechanism. This process involves calculating the relationship between each node... and its neighboring nodes Attention weights between To accomplish; Attention weight formula: ; in, and These are nodes eigenvectors, It is the LeakyReLU activation function. It is an exponential function. This indicates the concatenation of feature vectors between nodes. It is the weight matrix obtained through training. These are the learning parameters used to calculate attention. These are nodes Neighbor set; attention weight It is calculated based on the characteristics of switch nodes and is used to represent neighboring nodes. For nodes Updated influence; through attention mechanisms, nodes It will update its own characteristics based on the characteristics of its neighboring nodes; In graph neural networks, the updating of node features depends on the information of neighboring nodes. Through multi-layered attention mechanisms and information aggregation, the features of each node are gradually updated, ultimately yielding a new spatiotemporal feature representation for each node. After processing by the graph attention network, the feature vector of each node... These characteristics represent the state of a node in the network, reflecting the node's own properties and its relationship with neighboring nodes. Node feature update formula: ; in, It is a node In the The feature vector of the layer, It is a node In the The feature vector of the layer, These are the calculated attention weights; Finally, by summarizing the features of all nodes, the spatial state matrix is ​​obtained. : ; This matrix contains the spatiotemporal characteristic modeling representations of all nodes, serving as input for subsequent tasks such as traffic prediction and scheduling optimization; Predicting future traffic trends based on the Informer model To enhance model prediction results, spatial characteristics are used; first, the spatial state matrix is ​​received. and multi-step prediction sequences Next, the multi-step prediction sequence is subjected to decay-weighted aggregation along the time dimension to generate a time feature vector. : ; in These are weighting coefficients; Next, adaptive weights are applied to each aggregated business component. : ; Then, the temporal and spatial features are scale-normalized and aligned, and the aligned features are fused using a multilayer perceptron (MLP) to output the spatiotemporal fused features. This serves as input for subsequent ATS parameter optimization.

[0012] The present invention is further configured such that steps 3 and 4 are performed according to the following steps: First, based on the obtained spatiotemporal feature vectors Using an offline-trained DDQN network, the ATS parameter is implemented: token bucket capacity. Filling rate and gating list To optimize and update the system, offline training is adopted to ensure real-time performance. After training, a state-action mapping table is generated, which is used in the subsequent online phase to obtain the optimal parameters for the current scheduling cycle by looking up the table in milliseconds. At every moment The capacity of the token bucket Based on a fixed filling rate Update; the formula for calculating the token bucket capacity is: ; in, This is the maximum capacity of the token bucket. For a moment The number of tokens, The token bucket filling rate. For a moment The token consumption is directly proportional to the amount of data sent. It is the update interval; ATS controls the transmission time of each data frame, i.e., the time when a pass / fail condition is met; the earliest transmission time of each data packet is calculated based on the current token bucket size; a data frame is defined as... Then the qualified sending time The calculation formula is: ; in, For frames The earliest time of sending; The size of the frame; This represents the token bucket filling rate; The arrival time of the frame; When the token bucket overflows, the tokens exceeding the capacity are discarded, resulting in packet loss. The overflow control formula is: ; Among them, the number of overflowing tokens The number of tokens in the bucket Subtract maximum capacity Excessive amounts will result in packet loss; Define the number of hops the path traverses as port speed The token bucket limits the burst capacity to The frame size is Storage / propagation latency Serialization latency is , For a fixed delay, To delay queuing; Single-hop serialization + fixed overhead: , ; After plastic surgery, the upper limit of the queue: ; End-to-end upper bound: ; And constrain the end-to-end upper bound of key business operations. ,in This is the time delay threshold; In DDQN, state The spatial state matrix is ​​the input received by the model, representing the current spatiotemporal state of the network; therefore, in the above steps, the spatial state matrix... and traffic prediction results Feature vectors are formed by splicing and combining them. To serve as the input to the reinforcement learning model; define the spatiotemporal feature vector. for: ; in, It is the spatial state matrix of the network, including topology information, bandwidth, latency, etc. The traffic prediction results include traffic predictions for multiple future time steps; the spatiotemporal feature vectors are... As a state in a reinforcement learning model This state includes the network's real-time topology and expected changes in traffic; each state The data is then fed into DDQN for subsequent ATS parameter optimization. In reinforcement learning, actions This refers to the ATS parameters selected by the system, namely the token bucket capacity C, filling rate R, and gating list G used to control traffic transmission during network scheduling; these actions are the decision variables for the ATS parameters, and the optimization objective is to maximize the reward by configuring these parameters; in DDQN, the selection of actions is determined by the current state. It consists of two parts: action selection strategy and action selection strategy. At each time step, DDQN predicts... To select the optimal ATS parameters : ; in, yes The value represents the quality of a given state and action; These are the network parameters, which are updated through learning; reward function It is used to measure the degree of performance improvement of the network after selecting a certain action; in ATS parameter optimization, the reward function should be able to reflect the impact of traffic scheduling on performance indicators such as latency, packet loss rate, and jitter; the reward function is expressed as: ; in, It is a reward function. These are weighting coefficients. It is the average latency. This indicates end-to-end jitter. Indicates packet loss rate, This refers to the maximum latency exceeding the limit; the reward function is dynamically updated according to the ATS parameter configuration to optimize latency and bandwidth. The goal of DDQN is to minimize the current Value and Future The error between values ​​is updated via gradient descent. Value network; objective function Defined as: ; in It is a discount factor, representing the impact on future rewards; The maximum value of the next state computed by the target network. Value; and It is the TD error, used to update the parameters of the current network; pass Value updates are performed, and the DDQN model gradually optimizes its ATS parameter configuration to maximize traffic scheduling performance while minimizing losses such as latency and packet loss. After offline training is completed, a state-action mapping table is generated, and online decisions are made quickly by looking up the table. Within each scheduling cycle, the system determines the action based on the current state. Find the optimal ATS parameter configuration It is used to smooth burst traffic, allocate bandwidth and reduce latency jitter, and perform real-time traffic scheduling. Set the prediction window (in For the scheduling period, The optimal asynchronous flow shaper (ATS) parameter sequence, covering multiple short periods, is pre-calculated within the prediction window. ; During the online execution phase, the corresponding parameters are directly looked up in the table within each short cycle, avoiding the impact of online model inference delay on real-time performance.

[0013] The present invention is further configured such that step 5 is performed according to the following steps: Based on an asynchronous scheduling mechanism, a burst detection module is set up to avoid consequences such as traffic surges, link anomalies, or prediction mismatches. When the trigger conditions are met, it switches to preset conservative traffic shaping parameters. The detection content includes: Flow rate change detection : ; Prediction bias detection : ; Queues and Link Thresholds: in, , , and All are thresholds. For the rate of change of flow, To predict bias, For queues, For link threshold utilization, It is a real traffic sequence. It is a predicted flow sequence; When the detection finds that any of the above conditions exceed the threshold for a continuous period of time, that is... If so, redundant scheduling protection will be triggered; When anomalies are identified through the burst detection module, i.e. At that time, it will switch to the preset conservative flow shaping parameters. Conservative fill rate Conservative token bucket capacity and conservative gate list Assume the critical flow average to bottom rate is The suddenness coefficient is Maximum frame length Allow maximum in-film wait Then calculate the conservative parameters: ; in, It is the token bucket capacity, It is the fill rate, and G is the gated list; ; in, It is a conservative flow shaping parameter. This indicates that an emergency has occurred; when When the next cycle is in progress, conservative flow shaping parameters are used. ; When continuous If no anomaly is triggered within a certain number of cycles, the parameters will be gradually restored to the optimized parameters in steps of Δ, where Δ is the preset backoff step size.

[0014] An electronic device includes a memory and a processor, the memory being used to store a program that enables the processor to perform the methods described above, and the processor being configured to execute the program stored in the memory.

[0015] A computer-readable storage medium storing a computer program, characterized in that the computer program, when run by a processor, executes the steps of the above-described method.

[0016] The beneficial effects of this invention are as follows: Spatiotemporal feature fusion improves prediction accuracy: This invention combines spatial topology modeling of graph neural networks with long-period time series prediction of Informer to achieve deep fusion of spatial information such as network topology and link performance with temporal features of traffic trends, thereby providing a more comprehensive and accurate basis for ATS parameter optimization. The asynchronous prediction mechanism ensures millisecond-level real-time performance: This invention decouples complex prediction and optimization calculation from the millisecond-level scheduling cycle, and generates parameter sequences in advance within a longer prediction window, so that only the results need to be called directly within the scheduling cycle, which significantly reduces latency and meets the real-time requirements of critical business. Intelligent parameter optimization adapts to dynamic environments: This invention adopts a reinforcement learning strategy of offline training and table lookup to quickly obtain the optimal ATS parameter configuration based on the real-time network status, realize millisecond-level online decision-making, and maintain low latency and high reliability of critical business flows in dynamic scenarios such as business fluctuations and topology changes. Protection mechanisms for prediction failure and anomaly detection: This invention introduces a burst detection and redundant scheduling protection mechanism during parameter prediction and optimization. Once a prediction timeout or network status anomaly is detected, it can switch to a safe and conservative parameter configuration in a timely manner to ensure system stability and business continuity. Attached Figure Description

[0017] Fig. 1 This is a schematic diagram of the Informer model architecture for a deterministic latency guarantee method for network data transmission using Informer with enhanced spatial characteristics proposed in this invention. Fig. 2 This is a schematic diagram illustrating the workflow of a deterministic latency guarantee method for network data transmission using the spatial characteristic enhancement Informer proposed in this invention. Detailed Implementation

[0018] The technical solution of this patent will be further described in detail below with reference to specific embodiments.

[0019] The embodiments of this patent are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain this patent, and should not be construed as limiting this patent.

[0020] Reference Figs. 1-2 A deterministic latency guarantee method for network data transmission in Informer with enhanced spatial characteristics includes the following steps: Step 1: Informer Time Series Prediction Step 1.1: Constructing the input sequence First, the input data for this model consists of historical time-series data of multiple key business traffic types, such as GOOSE, SV, and MMS, i.e., a set of key business flows. These traffic data represent the transmission status of different service types in the network, including their temporal variation characteristics.

[0021] Input data: Time series data It includes multi-dimensional time-series data of historical traffic.

[0022] in, It is a fixed time step Observational data of GOOSE, SV, and MMS flows under the specified conditions. During the sampling period... Downsampling each service flow to obtain past The flow sequence at each time point: Among them, business traffic types Multiple service flows are concatenated into an input tensor: M=| | in This refers to the total number of business traffic types.

[0023] Step 1.2, Traffic Prediction Model Received via Informer model The Informer model encodes data to predict traffic flow for multiple scheduling cycles in a coal mine time-sensitive network. It employs a multi-head probabilistic sparse self-attention mechanism to uncover potential relationships between vectors and provides predictions for long-term series in a single step, effectively addressing the problem of increased errors caused by insufficient prediction of long-term series by traditional neural networks.

[0024] Step 1.2.1, Encoder Structure The Informer model primarily consists of an encoder and a decoder. Data first enters the encoder's embedding layer, which contains three encoding modules: an embedding encoding module, a position encoding module, and a timestamp encoding module. Each module encodes the input data, and the encoded data are then concatenated to obtain the encoded feature matrix. .

[0025] Next, the encoded feature matrix Processed into 3 matrices of the same size, namely the query matrix. Key matrix Value matrix ( They are respectively Sequence length, (for sequence feature dimensions) In the formula They are respectively The weight matrix.

[0026] Through the key matrix Random sampling, obtaining the sampled At this point, query matrix The i-th row The corresponding sparsity metric is in It is a sparsity metric used to determine the importance of the query vector in attention calculation; It is a query matrix The i-th row, for The OK, For sequence feature dimensions, It is the sequence length of the sampled key matrix. It takes the maximum value of all scores in the sample key set.

[0027] Select the corresponding sparsity metric with a higher value To form a new query matrix Calculate probabilistic sparse self-attention: in It is the dimension of the query vector. The transpose of the key matrix determines the scale of attention computation, and the attention weights are normalized through the Softmax operation.

[0028] To enhance the model's ability to capture long-term dependencies in long-sequence data while effectively reducing computational complexity, multi-head attention mechanisms concatenate the outputs of multiple different attention heads, integrating multi-dimensional traffic information: in It is the overall output of the multi-head attention mechanism. It is the output of multiple heads, representing the focus on different features; For splicing operations, It is a linear transformation matrix that generates the final prediction result.

[0029] To ensure computational efficiency, interlayer distillation is employed, extracting core features from each layer and using these as a foundation to construct a more efficient and focused self-attention representation in the next layer. The encoder distillation layer receives the first... The output of the layer after multi-head probabilistic sparse self-attention operation contains one-dimensional convolution and max pooling operations to achieve dimensionality compression. From the... Feature extraction from layer 1 is passed to layer 2. The layer operation process is as follows in: For the first time after distillation layer The feature matrix at time step; It is the max pooling function; for Activation function; It is a one-dimensional convolutional filter; For the first layer Time-time feature matrix; This represents the multi-head probability sparse self-attention operation.

[0030] Step 1.2.2, Decoder Structure The Informer model decoder consists of two multi-head self-attention layers. By introducing a generative decoder and a masking mechanism, the traffic prediction result depends only on the data preceding the sequence to be predicted. Under this mechanism, the data of the sequence to be predicted is masked (set to 0), ensuring that the model only focuses on the known data portion. The input to the decoder is the concatenation of the temporal feature matrix from the encoder and the sequence to be predicted.

[0031] The encoder and decoder interact via a cross-attention mechanism. The decoder utilizes the global feature information extracted by the encoder, combined with the contextual dependencies of the target sequence, to construct a feature matrix. This feature matrix is ​​then input into the fully connected layer to output the predicted traffic value. The output predicted sequence is... in This refers to the total number of business traffic types. The desired time step is determined by the prediction, so asynchronous prediction is used to ensure real-time performance.

[0032] Step 1.3, Loss Function To ensure the model can handle different types of traffic simultaneously, the loss function needs to consider the weight of each type of traffic and calculate the prediction error using mean squared error. Specifically, the loss function... By calculating each type of business traffic ( The mean squared error between the predicted and actual values ​​is calculated, then a weighted average is taken, and finally the errors for all traffic types are summed. The loss function is calculated using the following formula: in It is the traffic of each type of business. The weighting coefficients are typically assigned values ​​based on the importance of different traffic types. ; It predicts the time step. It is the model's prediction of the first Time of the first Predicted values ​​for traffic types, It is the actual first model Time of the first The actual value of traffic.

[0033] Step 2: Spatial Characteristic Model of TSN Network Step 2.1: Constructing the network topology and node characteristics First, a graph neural network is used to model the switch topology in a time-sensitive network (TSN), aiming to improve upon the shortcomings of the Informer class, which relies solely on time features and ignores network topology information, thereby enhancing spatial characteristics. Switches in the network are treated as nodes. A link is considered as an edge between nodes. Network topology diagram Define the characteristics of each node, including link bandwidth. Utilization rate Delay Bit error rate and node degree Parameters such as these. For each switch node. Its eigenvectors include: in, This indicates the link bandwidth, reflecting the link's transmission capacity; Indicates link utilization; This represents the link delay between node i and its neighbor node j; Bit error rate (BER) represents the probability that a data packet will encounter an error during transmission; while Represents a node The degree of these nodes is used as input to the graph neural network for further processing.

[0034] Step 2.2: Calculate the attention weights between nodes. Next, we extract network features using the Graph Attention (GAT) mechanism in graph neural networks. In a graph attention network, the relationships between nodes, i.e., the degree of influence each node has on its neighbors, are calculated through the attention mechanism. This process involves calculating the relationship between each node... and its neighboring nodes Attention weights between It was accomplished through this process.

[0035] Attention weight formula: in, and These are nodes eigenvectors, It is the LeakyReLU activation function. It is an exponential function. This indicates the concatenation of feature vectors between nodes. It is the weight matrix obtained through training. These are the learning parameters used to calculate attention. These are nodes Neighbor set. Attention weight. It is calculated based on the characteristics of switch nodes and is used to represent neighboring nodes. For nodes Updated influence. Through attention mechanisms, nodes... It will update its own characteristics based on the characteristics of its neighboring nodes.

[0036] Step 2.3: Generate the spatial state matrix In graph neural networks, the updating of node features depends on information from neighboring nodes. Through multi-layered attention mechanisms and information aggregation, the features of each node are progressively updated, ultimately yielding a new spatiotemporal feature representation for each node. After processing by the graph attention network, the feature vector of each node... These characteristics represent the state of a node in the network, reflecting the node's own properties and its relationship with neighboring nodes.

[0037] Node feature update formula: in, It is a node In the The feature vector of the layer, It is a node In the The feature vector of the layer, These are the calculated attention weights.

[0038] Finally, by summarizing the features of all nodes, the spatial state matrix is ​​obtained. : This matrix contains the spatiotemporal characteristics modeled for all nodes, and is subsequently used as input for tasks such as traffic prediction and scheduling optimization.

[0039] Step 2.4, Enhancement of Spatial Characteristics Predicting future traffic trends based on the Informer model Then, spatial characteristics are used to enhance the prediction results of the Informer model. First, the spatial state matrix is ​​received. and multi-step prediction sequences Next, the multi-step prediction sequence is aggregated with decay weights along the time dimension to generate a time feature vector. : in These are weighting coefficients.

[0040] Next, adaptive weights are applied to each aggregated business component. : Then, the temporal and spatial features are scale-normalized and aligned, and the aligned features are fused using a multilayer perceptron (MLP) to output the spatiotemporal fused features. This serves as input for subsequent ATS parameter optimization. A prediction-fusing method is employed to enhance spatial characteristics.

[0041] Step 3: Intelligent Optimization of ATS Parameters Based on the obtained spatiotemporal feature vectors Using an offline-trained DDQN network, the ATS parameter is implemented: token bucket capacity. Filling rate and gating list To optimize and update the system, offline training is adopted to ensure real-time performance. After training, a state-action mapping table is generated, which is used in the subsequent online phase to obtain the optimal parameters for the current scheduling cycle in a millisecond-level table lookup manner.

[0042] Step 3.1, Token Bucket and Qualification Occurrence Time Model (1) Token bucket capacity: At every moment The capacity of the token bucket Based on a fixed filling rate Update. The formula for calculating the token bucket capacity is: in This is the maximum capacity of the token bucket. For a moment The number of tokens, The token bucket filling rate. For a moment The token consumption is directly proportional to the amount of data sent. It is the update interval.

[0043] (2) Qualified transmission time ATS controls the transmission time of each data frame, i.e., the time when a pass / fail condition is met. The earliest transmission time for each data packet is calculated based on the current token bucket size. A data frame is defined as... Then the qualified sending time The calculation formula is: in, For frames The earliest time it can be sent; The size of the frame; This represents the token bucket filling rate; This represents the arrival time of the frame.

[0044] (3) Token bucket overflow handling When the token bucket overflows, tokens exceeding the capacity are discarded, resulting in packet loss. The overflow control formula is as follows: The number of tokens that overflowed. The number of tokens in the bucket Subtract maximum capacity Excessive amounts will result in packet loss.

[0045] Step 3.2, End-to-End Delay and Constraints Define the number of hops the path traverses as port speed The token bucket limits the burst capacity to The frame size is Storage / propagation latency Serialization latency is , For a fixed delay, To delay queuing time. Single-hop serialization + fixed overhead: , .

[0046] After plastic surgery, the upper limit of the queue: .

[0047] End-to-end upper bound: And constrain the end-to-end upper bound of key business operations. ,in This is the time delay threshold.

[0048] Step 3.3: Representation of State, Action, and Reward Function In DDQN, state This is the input received by the model, used to represent the current spatiotemporal state of the network. Therefore, in the above steps, the spatial state matrix... and traffic prediction results Feature vectors are formed by splicing and combining them. As input to the reinforcement learning model, define the spatiotemporal feature vector. for: in It is the spatial state matrix of the network, including topology information, bandwidth, latency, etc. The traffic prediction results include traffic forecasts for multiple future time steps. The spatiotemporal feature vectors... As a state in a reinforcement learning model This state includes the network's real-time topology and expected changes in traffic. Each state The data is then input into DDQN for subsequent ATS parameter optimization.

[0049] In reinforcement learning, actions This refers to the ATS parameters selected by the system, namely the token bucket capacity C, filling rate R, and gating list G used to control traffic transmission during network scheduling. These actions are the decision variables for the ATS parameters, and the optimization objective is to maximize the reward by configuring these parameters. In DDQN, the selection of actions is determined by the current state. It consists of two parts: action selection strategy and action selection strategy.

[0050] At each time step, DDQN predicts... To select the optimal ATS parameters : in yes The value represents the quality of a given state and action. These are the network parameters, which are updated through learning.

[0051] reward function It is used to measure the performance improvement of the network after selecting a certain action. In ATS parameter optimization, the reward function should reflect the impact of traffic scheduling on performance metrics such as latency, packet loss rate, and jitter. The reward function is expressed as follows: in It is a reward function. These are weighting coefficients. It is the average latency. This indicates end-to-end jitter. Indicates packet loss rate, This refers to the maximum latency exceeding the limit. The reward function is dynamically updated based on the ATS parameter configuration to optimize latency and bandwidth.

[0052] Step 3.4: DDQN Objectives and Training Process The goal of DDQN is to minimize the current Value and Future The error between values ​​is updated via gradient descent. Value network. Objective function. Defined as: in It is a discount factor, representing the impact of future rewards. The maximum value of the next state computed by the target network. Value. It is the TD error, used to update the parameters of the current network.

[0053] pass With value updates, the DDQN model gradually optimizes its ATS parameter configuration to maximize traffic scheduling performance while minimizing losses such as latency and packet loss.

[0054] Step 3.5: Online Inquiry and Decision Making After offline training is complete, a state-action mapping table is generated, enabling rapid online decision-making through table lookup. Within each scheduling cycle, the system determines the action based on the current state. Find the optimal ATS parameter configuration It is used to smooth burst traffic, allocate bandwidth and reduce latency jitter, and perform real-time traffic scheduling.

[0055] Step 4: Decoupling Asynchronous Prediction and Scheduling Step 4.1, Prediction Calculation: Set the prediction window (in For the scheduling period, The optimal asynchronous flow shaper (ATS) parameter sequence, covering multiple short periods, is pre-calculated within the prediction window. ; Step 4.2, Short-cycle scheduling During the online execution phase, the corresponding parameters are directly looked up in the table within each short cycle, avoiding the impact of online model inference delay on real-time performance.

[0056] Step 5: Emergency Detection and Protection Step 5.1, Burst Detection Based on an asynchronous scheduling mechanism, a burst detection module is set up to avoid consequences such as traffic surges, link anomalies, or prediction mismatches. When the trigger conditions are met, it switches to preset conservative traffic shaping parameters. The detection content includes: Flow rate change detection Prediction bias detection Queues and Link Thresholds in , , and All are thresholds. For the rate of change of flow, To predict bias, For queues, For link threshold utilization, It is a real traffic sequence. It is a predicted flow sequence.

[0057] When the detection finds that any of the above conditions exceed the threshold for a continuous period of time, that is... If this occurs, redundant scheduling protection will be triggered.

[0058] Step 5.2, Redundancy Scheduling Protection and Recovery (1) Calculation of conservative parameters When anomalies are identified through the burst detection module, i.e. At that time, it will switch to the preset conservative flow shaping parameters. Conservative fill rate Conservative token bucket capacity and conservative gate list Let the average bottom rate of the critical flow be... The suddenness coefficient is Maximum frame length Allow maximum in-film wait Then calculate the conservative parameters: in It is the token bucket capacity, It is the fill rate, and G is the gated list.

[0059] Switch strategy (effective in the next cycle) in It is a conservative flow shaping parameter. This indicates that an emergency has occurred.

[0060] when When the next cycle is in progress, conservative flow shaping parameters are used. .

[0061] Gradual recovery: when continuous If no anomaly is triggered in a given cycle, the parameters will be gradually restored to the optimized parameters in steps of Δ, where Δ is the preset backoff step size.

[0062] In this embodiment, an electronic device includes a memory and a processor. The memory stores a program that supports the processor in executing the methods described above, and the processor is configured to execute the program stored in the memory.

[0063] In this embodiment, a computer-readable storage medium stores a computer program, which is executed by a processor to perform the steps of the above method.

[0064] The Informer model consists of an encoder, a decoder, and a fully connected output. The encoder takes historical multi-service traffic sequences as input, extracts long-period features through multi-head probabilistic sparse self-attention, and compresses the sequence length through inter-layer distillation. The decoder receives the masked sequence to be predicted, cross-attentions it with the encoder features, and outputs a prediction of the traffic in the next H steps through a fully connected layer, achieving a long-period result in a single forward pass, balancing accuracy and efficiency.

[0065] First, the time series of critical business traffic is predicted based on the Informer model, and long-term time features and future traffic trends are extracted through a self-attention mechanism. Next, a graph neural network is introduced to model the topology of switches in the TSN network, and the attention mechanism is used to quantify link bandwidth and latency characteristics to generate a spatial state matrix. The spatiotemporal features are then fused to generate optimal asynchronous traffic shaper parameters, used to smooth burst traffic, allocate bandwidth, and reduce latency jitter. Simultaneously, to ensure real-time performance, an asynchronous prediction and short-cycle scheduling mechanism is designed to decouple prediction calculation from short-cycle scheduling execution. During operation, burst detection and redundant scheduling protection modules reserve bandwidth and priority for critical business flows, ensuring deterministic latency of critical business data in extreme scenarios.

[0066] 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 deterministic delay guarantee method for network data transmission in a spatially enhanced Informer, comprising the following steps: Step 1: Based on the Informer model, perform traffic prediction for multiple types of key business traffic; Constructing multivariate time series data from historical traffic data Furthermore, the ProbSparse self-attention mechanism is used to extract long-term temporal features and predict traffic trends for multiple future scheduling cycles. ; Step 2: Construct a spatial characteristic model of the TSN network based on a graph neural network, treating the switches as a set of nodes. Links as a set of edges Node features Including link bandwidth Utilization rate Delay Bit error rate and node degree Through graph attention networks, the nodes are computed. Its neighboring nodes Attention weights are used to generate a spatial state matrix that reflects the real-time transmission capability of the network. Next, using the spatial state matrix Spatial feature enhancement is applied to the model in step 1, incorporating future traffic sequences. With spatial state matrix Feature fusion is performed to obtain the spatiotemporal feature vector for subsequent inputs. ; Step 3: Based on the obtained spatiotemporal feature vectors By utilizing an offline-trained network, and with the optimization objective of minimizing the end-to-end average latency, jitter, and packet loss rate of critical business operations, the optimal asynchronous traffic shaper token bucket capacity is generated. Filling rate and gating list After training is completed, a state-action mapping table is generated, which is used in the subsequent online phase to obtain the optimal parameters for the current scheduling cycle in a millisecond-level table lookup manner. Step 4: Set up an asynchronous scheduling mechanism to decouple prediction calculation from short-cycle scheduling execution, and pre-calculate the optimal asynchronous flow shaper parameter sequence covering multiple short cycles within the prediction window. It is used to smooth burst traffic, allocate bandwidth, and reduce latency jitter; During the online execution phase, the corresponding parameters are directly looked up and called from the table within each short cycle; Step 5: During operation, set up a burst detection algorithm to monitor for traffic spikes, link anomalies, or prediction mismatches in real time based on sliding window statistical calculations. When the trigger conditions are met, switch to the preset conservative traffic shaping parameters. Reserve bandwidth and priority for critical business flows until the prediction module recovers and then switch back to the optimized parameter configuration.

2. The method for guaranteeing deterministic latency of network data transmission in a spatially enhanced Informer according to claim 1, characterized in that, Step 1 is performed as follows: First, the input data for this model consists of historical time-series data of multiple key business traffic types, including GOOSE, SV, and MMS, i.e., a set of key business flows. These traffic data represent the transmission status of different service types in the network, including their temporal variation characteristics. Input data: Time series data This includes multi-dimensional time-series data of historical traffic; ； in, It is a fixed time step Observational data of GOOSE, SV, and MMS flows under the specified conditions; during the sampling period Downsampling each service flow to obtain past The flow sequence at each time point: ； Among them, business traffic types Multiple service flows are concatenated into an input tensor: ，M=| |； in, This refers to the total number of business traffic types; Received via Informer model The data is encoded to enable traffic prediction for multiple future scheduling cycles of time-sensitive networks in coal mines. The Informer model employs a multi-head probabilistic sparse self-attention mechanism to uncover potential relationships between vectors and provide prediction results for long-term series in one go.

3. The method for guaranteeing deterministic latency in network data transmission using a spatially enhanced Informer according to claim 2, characterized in that, The Informer model mainly consists of an encoder and a decoder. Data first enters the encoder's embedding layer, which contains three encoding modules: an embedding encoding module, a position encoding module, and a timestamp encoding module. Each module encodes the input data, and the encoded data are then concatenated to obtain the encoded feature matrix. ; Next, the encoded feature matrix Processed into 3 matrices of the same size, namely the query matrix. Key matrix Value matrix They are respectively Sequence length, For sequence feature dimensions: ； In the formula, They are respectively The weight matrix; Through the key matrix Random sampling, obtaining the sampled At this point, query matrix The i-th row The corresponding sparsity metric is: ； in, It is a sparsity metric used to determine the importance of the query vector in attention calculation; It is a query matrix The i-th row, for The OK, For sequence feature dimensions, It is the sequence length of the sampled key matrix. It takes the maximum value of all scores in the sample key set; Select the corresponding sparsity metric with a higher value To form a new query matrix Calculate probabilistic sparse self-attention: ； in, It is the dimension of the query vector. The transpose of the key matrix determines the scale of attention computation, and the attention weights are normalized through the Softmax operation. To enhance the model's ability to capture long-term dependencies in long-sequence data while effectively reducing computational complexity, multi-head attention mechanisms concatenate the outputs of multiple different attention heads, integrating multi-dimensional traffic information: ； in, It is the overall output of the multi-head attention mechanism. It is the output of multiple heads, representing the focus on different features; For splicing operations, It is a linear transformation matrix that generates the final prediction result; To ensure computational efficiency, interlayer distillation is used to extract core features from each layer, and these features are then used as a basis to build a more efficient and focused self-attention representation in the next layer; the encoder distillation layer receives the first... The output of the layer after multi-head probabilistic sparse self-attention operation; this layer contains one-dimensional convolution and max pooling operations to achieve dimensionality compression; from the th Feature extraction from layer 1 is passed to layer 2. The layer operation process is as follows: ； in, For the first time after distillation layer The feature matrix at each time step; It is the max pooling function; for Activation function; It is a one-dimensional convolutional filter; For the first layer Time-time feature matrix; This represents the multi-head probability sparse self-attention operation.

4. The method for guaranteeing deterministic latency of network data transmission in a spatially enhanced Informer according to claim 3, characterized in that, The Informer model decoder consists of two multi-head self-attention layers. By introducing a generative decoder and a masking mechanism, the traffic prediction result depends only on the data preceding the sequence to be predicted. Under this mechanism, the data of the sequence to be predicted is masked to ensure that the model only focuses on the known data. The input of the decoder is the concatenation of the temporal feature matrix from the encoder and the sequence to be predicted. The encoder and decoder exchange information through a cross-attention mechanism; The decoder utilizes the global feature information extracted by the encoder and combines it with the contextual dependencies of the target sequence to construct a feature matrix. This feature matrix is ​​then input into the fully connected layer to output the predicted flow rate. The output predicted sequence is: ； in, This refers to the total number of business traffic types. The desired time step is predicted, so asynchronous prediction is used to ensure real-time performance. To ensure the model can handle different types of traffic simultaneously, the loss function needs to consider the weight of each type of traffic and calculate the prediction error using mean squared error; specifically, the loss function... By calculating each type of business traffic ( The mean squared error between the predicted and actual values ​​is calculated, then a weighted average is taken, and finally the errors for all traffic types are summed; the loss function is calculated as follows: ； in, It is the traffic of each type of business. The weighting coefficients are typically assigned values ​​based on the importance of different traffic types. ; It predicts the time step. It is the model's prediction of the first Time of the first Predicted values ​​for traffic types, It is the actual first model Time of the first The actual value of traffic.

5. The method for guaranteeing deterministic latency in network data transmission using a spatially enhanced Informer according to claim 1, characterized in that, Step 2 is performed as follows: First, a graph neural network is used to model the switch topology in a Time-Sensitive Network (TSN); the switches in the network are treated as nodes. A link is considered as an edge between nodes. Network topology diagram Define the characteristics of each node, including link bandwidth. Utilization rate Delay Bit error rate and node degree Five parameters; for each switch node , Its eigenvectors include: ； in, This indicates the link bandwidth, reflecting the link's transmission capacity; Indicates link utilization; This represents the link delay between node i and its neighbor node j; Bit error rate (BER) represents the probability that a data packet will encounter an error during transmission; while Represents a node The degree of these nodes; these node features are used as input to the graph neural network for further processing; Next, network features are extracted using the Graph Attention Network (GAT) mechanism in graph neural networks. In a graph attention network, the relationships between nodes, i.e., the degree of influence each node has on its neighboring nodes, are calculated through the attention mechanism. This process involves calculating the relationship between each node... and its neighboring nodes Attention weights between To accomplish; Attention weight formula: ； in, and These are nodes eigenvectors, It is the LeakyReLU activation function. It is an exponential function. This indicates the concatenation of feature vectors between nodes. It is the weight matrix obtained through training. These are the learning parameters used to calculate attention. These are nodes Neighbor set; attention weight It is calculated based on the characteristics of switch nodes and is used to represent neighboring nodes. For nodes Updated influence; through attention mechanisms, nodes It will update its own characteristics based on the characteristics of its neighboring nodes; In graph neural networks, the updating of node features depends on the information of neighboring nodes. Through multi-layered attention mechanisms and information aggregation, the features of each node are gradually updated, ultimately yielding a new spatiotemporal feature representation for each node. After processing by the graph attention network, the feature vector of each node... These characteristics represent the state of a node in the network, reflecting the node's own properties and its relationship with neighboring nodes. Node feature update formula: ； in, It is a node In the The feature vector of the layer, It is a node In the The feature vector of the layer, These are the calculated attention weights; Finally, by summarizing the features of all nodes, the spatial state matrix is ​​obtained. : ； This matrix contains the spatiotemporal characteristic modeling representations of all nodes, serving as input for subsequent traffic prediction and scheduling optimization tasks; Predicting future traffic trends based on the Informer model To enhance model prediction results, spatial characteristics are used; first, the spatial state matrix is ​​received. and multi-step prediction sequences Next, the multi-step prediction sequence is subjected to decay-weighted aggregation along the time dimension to generate a time feature vector. : ； in These are weighting coefficients; Next, adaptive weights are applied to each aggregated business component. : ； Then, the temporal and spatial features are scale-normalized and aligned, and the aligned features are fused using a multilayer perceptron (MLP) to output the spatiotemporal fused features. This serves as input for subsequent ATS parameter optimization.

6. The method for guaranteeing deterministic latency in network data transmission using a spatially enhanced Informer according to claim 1, characterized in that, Steps 3 and 4 are performed as follows: First, based on the obtained spatiotemporal feature vectors Using an offline-trained DDQN network, the ATS parameter is implemented: token bucket capacity. Filling rate and gating list To optimize and update the system, offline training is adopted to ensure real-time performance. After training, a state-action mapping table is generated, which is used in the subsequent online phase to obtain the optimal parameters for the current scheduling cycle by looking up the table in milliseconds. At every moment The capacity of the token bucket Based on a fixed filling rate Update; the formula for calculating the token bucket capacity is: ； in, This is the maximum capacity of the token bucket. For a moment The number of tokens, The token bucket filling rate. For a moment The token consumption is directly proportional to the amount of data sent. It is the update interval; ATS controls the transmission time of each data frame, i.e., the time when a pass / fail condition is met; the earliest transmission time of each data packet is calculated based on the current token bucket size; a data frame is defined as... Then the qualified sending time The calculation formula is: ； in, For frames The earliest time of sending; The size of the frame; This represents the token bucket filling rate; The arrival time of the frame; When the token bucket overflows, the tokens exceeding the capacity are discarded, resulting in packet loss. The overflow control formula is: ； Among them, the number of overflowing tokens The number of tokens in the bucket Subtract maximum capacity Excessive amounts will result in packet loss; Define the number of hops the path traverses as port speed The token bucket limits the burst capacity to The frame size is Storage / propagation latency Serialization latency is , For a fixed delay, To delay queuing; Single-hop serialization + fixed overhead: , ; After plastic surgery, the upper limit of the queue: ; End-to-end upper bound: ; And constrain the end-to-end upper bound of key business operations. ,in This is the time delay threshold; In DDQN, state The spatial state matrix is ​​the input received by the model, representing the current spatiotemporal state of the network; therefore, in the above steps, the spatial state matrix... and traffic prediction results Feature vectors are formed by splicing and combining them. To serve as the input to the reinforcement learning model; define the spatiotemporal feature vector. for: ； in, It is the spatial state matrix of the network, including topology information, bandwidth, and latency information; The traffic prediction results include traffic predictions for multiple future time steps; the spatiotemporal feature vectors are... As a state in a reinforcement learning model This state includes the network's real-time topology and expected changes in traffic; each state The data is then fed into DDQN for subsequent ATS parameter optimization. In reinforcement learning, actions This refers to the ATS parameters selected by the system, namely the token bucket capacity C, filling rate R, and gating list G used to control traffic transmission during network scheduling; these actions are the decision variables for the ATS parameters, and the optimization objective is to maximize the reward by configuring these parameters; in DDQN, the selection of actions is determined by the current state. It consists of two parts: action selection strategy and action selection strategy. At each time step, DDQN predicts... To select the optimal ATS parameters : ； in, yes The value represents the quality of a given state and action; These are the network parameters, which are updated through learning; reward function It is used to measure the degree of performance improvement of the network after selecting a certain action; in ATS parameter optimization, the reward function should be able to reflect the impact of traffic scheduling on performance indicators such as latency, packet loss rate, and jitter; the reward function is expressed as: ； in, It is a reward function. These are weighting coefficients. It is the average latency. This indicates end-to-end jitter. Indicates packet loss rate, This refers to the maximum latency exceeding the limit; the reward function is dynamically updated according to the ATS parameter configuration to optimize latency and bandwidth. The goal of DDQN is to minimize the current Value and Future The error between values ​​is updated via gradient descent. Value network; objective function Defined as: ； in It is a discount factor, representing the impact on future rewards; The maximum value of the next state computed by the target network. Value; and It is the TD error, used to update the parameters of the current network; pass Value updates are performed, and the DDQN model gradually optimizes its ATS parameter configuration to maximize traffic scheduling performance while minimizing latency and packet loss. After offline training is completed, a state-action mapping table is generated, and online decisions are made quickly by looking up the table. Within each scheduling cycle, the system determines the action based on the current state. Find the optimal ATS parameter configuration It is used to smooth burst traffic, allocate bandwidth and reduce latency jitter, and perform real-time traffic scheduling. Set the prediction window (in For the scheduling period, The optimal asynchronous flow shaper (ATS) parameter sequence, covering multiple short periods, is pre-calculated within the prediction window. ; During the online execution phase, the corresponding parameters are directly looked up in the table within each short cycle, avoiding the impact of online model inference delay on real-time performance.

7. The method for guaranteeing deterministic latency in network data transmission using a spatially enhanced Informer according to claim 1, characterized in that, Step 5 is performed as follows: Based on an asynchronous scheduling mechanism, a burst detection module is set up to avoid the consequences of traffic surges, link anomalies, or prediction mismatches. When the trigger conditions are met, it switches to preset conservative traffic shaping parameters. The detection content includes: Flow rate change detection : ； Prediction bias detection : ； Queues and Link Thresholds: in, , , and All are thresholds. For the rate of change of flow, To predict bias, For queues, For link threshold utilization, It is a real traffic sequence. It is a predicted flow sequence; When the detection finds that any of the above conditions exceed the threshold for a continuous period of time, that is... If so, redundant scheduling protection will be triggered; When anomalies are identified through the burst detection module, i.e. At that time, it will switch to the preset conservative flow shaping parameters. Conservative fill rate Conservative token bucket capacity and conservative gating list Assume the critical flow average to bottom rate is The suddenness coefficient is Maximum frame length Allow maximum in-film wait Then calculate the conservative parameters: ； in, It is the token bucket capacity, It is the fill rate, and G is the gated list; ； in, It is a conservative flow shaping parameter. This indicates that an emergency has occurred; when When the next cycle is in progress, conservative flow shaping parameters are used. ; When continuous If no anomaly is triggered within a certain number of cycles, the parameters will be gradually restored to the optimized parameters in steps of Δ, where Δ is the preset backoff step size.

8. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store a program that supports the processor in executing any of the deterministic delay guarantee methods of claims 1-7, and the processor is configured to execute the program stored in the memory.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, The computer program, when executed by the processor, performs the steps of any of the deterministic delay guarantee methods described in claims 1-7.

Images