Efficient routing methods for ultra-high bandwidth services in time-sensitive networks

Abstract

This invention discloses an efficient routing method for ultra-high bandwidth (UBF) service flows in Time-Sensitive Networks (TSNs), aiming to solve problems such as load balancing of UBF and improve network performance. The main steps include distinguishing UBF from other data flows to meet their respective needs; then, employing a multi-objective optimization efficient routing algorithm to select the optimal path for transmission for each service flow. This invention can improve the operational efficiency and performance of TSNs, especially in scenarios involving mixed transmission of ultra-high bandwidth and regular service flows, providing more reliable services for real-time applications.

Description

Technical Field

[0001] This invention relates to the field of time-sensitive networking technology for Internet communication networks, and more specifically, to an efficient routing method for ultra-high bandwidth service flows in time-sensitive networks. Background Technology

[0002] With the development of AR / VR technology, the high-speed interaction and precise control of immersive video transmission networks require data synchronization, deterministic latency, and ultra-high bandwidth. Simultaneously, the more complex communication modes of large-scale cluster computing technology place extremely high demands on network transmission capabilities. The emergence of immersive video transmission networks and the development of large-scale cluster computing technology have further propelled the expansion of the application scope of Time-Sensitive Networking (TSN). However, the communication data model of large-scale distributed intelligent computing systems exhibits characteristics of multi-point communication and "elephant streams." This change in the extreme business scenarios with high bandwidth requirements places higher demands on network performance in terms of real-time performance, reliability, and security. To address these challenges, deterministic real-time communication has been introduced into fields such as aerospace and industrial automation to ensure the performance of time-critical flows and the security of the system. TSN provides deterministic low-latency transmission services for Layer 2 networks. Specifically, the 802.1 TSN standard is designed to support the joint transmission of multiple types of traffic within the same network. However, the large volume of periodic control flows, scheduled ultra-high bandwidth flows, and ordinary IT flows continuously generated by IoT devices pose new challenges to the TSN network's ability to manage, route, and accommodate data flows.

[0003] There are significant differences in characteristics between Regular Business Flow (RSF) and Ultra-High Bandwidth Streaming (UBF). RSF is a major traffic category frequently optimized in TSN networks, including Time-Triggered (TT) streams, Audio Video Bridging (AVB) streams, and Best Effort (BE) streams. Typically, TT traffic is generated by control or command applications with strict real-time requirements and necessitates stringent latency constraints. AVB traffic originates from applications with less stringent real-time requirements. BE traffic is a basic traffic type without specific low latency or bandwidth requirements. UBF refers to traffic requiring extremely high bandwidth and deterministic transmission in the network. It is commonly used in scenarios such as big data processing in cloud computing, real-time video streaming, and ultra-low latency communication in telecommunications networks. Its transmission can lead to network congestion and bandwidth bottlenecks, resulting in load imbalances and impacting the performance of real-time applications. Therefore, it is necessary to distinguish UBF from other data streams to meet their respective requirements. Compared to traditional networks, TSN not only focuses on traffic transmission timing but also supports the transmission of critical traffic at the network layer by optimizing paths. Effective routing strategies can reduce transmission latency, improve resource utilization efficiency, and ensure the quality of service for critical traffic. Therefore, in a feasible and practical TSN system, it is crucial to simultaneously support the hybrid transmission of UBF and RSF.

[0004] Route optimization is crucial for ensuring UBF (Unified Flow) transmission quality. In TSN (Transportation Service Network) systems, heuristic solutions are typically used to route service flows along different paths to achieve spatial isolation. However, this heuristic approach often results in longer solution times and only considers RSF (Resolved Flow), neglecting the requirements and impact of UBF. Driven by services such as immersive video transmission networks and large-scale cluster computing technologies, UBF with ultra-high bandwidth and deterministic latency has become the most typical characteristic of new services. The significant differences in their traffic characteristics can lead to network congestion and bandwidth bottlenecks. Therefore, it is necessary to comprehensively consider the routing problem of mixed RSF and UBF deployments within the deterministic domain. However, due to the mutual influence between various flows in TSN and the existence of various constraints, designing an efficient routing solution is very challenging.

[0005] Existing routing methods for TSN multi-target mixed traffic flows and their limitations:

[0006] (1) For new scenarios with ultra-high bandwidth requirements, improving network operating efficiency in complex TSN networks with multiple paths and multiple inputs / outputs has become a challenging problem. To date, the design scheme for joint path optimization of regular service flows and ultra-high bandwidth flows in the scheduling and routing mechanisms of the TSN standard has not been effectively solved. Therefore, further research is needed on existing time-sensitive network routing management and communication scheduling methods. These issues directly relate to and affect the network's real-time performance and reliability.

[0007] (2) In TSN local area scenarios, existing solutions fail to effectively address the load balancing problem when UBF (Unified Flow Filter) exists. Considering the significant differences in traffic characteristics between UBF and RSF (Restricted Flow Filter), traditional load balancing routing algorithms fail to distinguish between the needs of UBF and RSF, ignoring the specific routing requirements of UBF, leading to conflicts between service flows, increasing the latency of time-sensitive flows, and thus violating strict flow deadlines. Furthermore, existing methods limit the diversity of solutions and fail to fully utilize the characteristics of TSN.

[0008] (3) Solving routing and scheduling problems of various flows based on machine learning or deep learning aims to minimize the end-to-end latency of RSF and construct an objective function consisting of a weighted sum of three indicators. However, the linear weighting method is difficult to accurately set the weight vector to obtain the Pareto optimal solution. Furthermore, traditional multi-objective optimization algorithms, such as Deep Reinforcement Learning (DRL), are prone to getting trapped in local optima and are unable to fully explore the solution space to find the global optimal solution. The Non-dominated sorting genetic algorithm (NSGA-II) takes a long time to find the optimal solution and has limited local optimization capabilities. Summary of the Invention

[0009] This invention addresses the routing optimization problem in time-sensitive networks (TSNs) with a mix of high-bandwidth and regular traffic flows. It provides an efficient routing method for high-bandwidth traffic flows in TSNs to solve problems such as high latency and unbalanced link load in the network.

[0010] The technical solution to achieve the purpose of this invention is: an efficient routing method for ultra-high bandwidth service flows in time-sensitive networks, comprising the following steps:

[0011] S1: The centralized user configuration (CUC) initiates a request to the network central controller (CNC) through the user network interface (UNI) to retrieve the network physical topology information of the TSN network; the CNC discovers the TSN network topology, abstracts it into a directed network graph, and returns the result to the CUC;

[0012] S2: The terminal system sends TSN connection requests and resource requirements to the CUC through the user configuration protocol, including information such as the communication terminal, the period T, size and deadline of the service flow, in order to obtain all traffic information in the network. The CUC then transmits these connection requests and resource requirements to the CNC through UNI.

[0013] S3: The CNC performs service modeling at the switch based on the service flow attributes. First, the controller CNC divides these service flows into TT flow, AVB flow, and BE flow based on the QoS latency requirements of different flows. Then, when the bandwidth of the data flow reaches a certain threshold of the link bandwidth, it will be identified and classified as UBF. At the same time, PCP and VID are assigned to each service flow and mapped to different priority queues of the switch.

[0014] S4: Obtain basic information about the entire network, including the TSN network topology map and data flow information set, and establish a multi-objective optimization mathematical model that considers multiple routing constraints based on this basic information;

[0015] S5: CNC performs routing calculations based on the network physical topology, network requirements, and a multi-objective optimization mathematical model containing multiple routing constraints to obtain relevant path information that meets connectivity requirements and is feasible. The first stage of the routing calculation is to use Transformer-based deep reinforcement learning T-DRL to obtain a feasible initial solution as an alternative path.

[0016] S6: The second stage of route calculation is performed. The CNC optimizes the candidate paths selected in step S5 using the Non-Dominated Sorting Genetic Algorithm II (NSGA-II) to select the preferred path from the sender to the receiver that meets the service flow requirements.

[0017] S7: CNC traverses to see if there are any uncalculated paths for the sending and receiving ends. If so, return to steps S3-S6; otherwise, proceed to step S8.

[0018] S8: The CNC encapsulates the calculated path results into a transmission table, configures it to the TSN switch, and then sends it to the TSN terminal device through the CNC.

[0019] Further, step S1 specifically includes: before the network starts operating, the CUC initiates a request to the CNC to retrieve the network physical topology through the UNI; the CNC discovers the TSN network topology information and node information according to LLDP (Link Layer Discovery Protocol), and models it into a directed TSN graph G = (V, E) using network modeling methods, and returns the result to the CUC; the specific construction process of the network model is as follows:

[0020] The TSN network topology is modeled as a directed graph G = (V, E), where V ≡ (v1, v2, ..., v...). n+λ () represents a set of n switch (SW) nodes and a set of λ terminal device (ES) nodes, where E = {e ij |i,j∈n,i≠j} is the set of physical links. TSN supports full-duplex, and each full-duplex link is treated as two separate directed links e. ij =(v i ,v j 0 and e ji =(v j ,v i Let A∈R n×n Let e ​​represent an adjacency matrix, a matrix consisting of elements 0 and 1, where e ∈ [0, 1], and e ∈ [0, 1]. ij ∈E, A=1, if A = 0; each link e ij Defined by triples <lb ij ,ld ij ,lt ij >, among which It is link e ij bandwidth capacity, The link delay, expressed in μsec, is determined by the transmission delay pd of i. ij Queuing delay qd ij and link (v i ,v j The propagation delay td ij composition, To pass through link (v i ,v j The number of TSN streams.

[0021] Further, step S2 includes: the terminal system configuring the communication terminal (src) via a user configuration protocol. k ,dst k ), base period τ of business flow k Deadline θ k Bandwidth size B k Priority ρ k Population size (Size) and maximum number of iterations (ζ) in NSGA-II NSGA-II Maximum number of iterations ζ in T-DRL T-DRL The system sends information to the CUC to retrieve all information in the network. The CUC then forwards these communication requests to the CNC via the UNI, as follows:

[0022] Each communication task is abstracted as a service flow, and these service flows are then defined by tuples f. k ≡(src k ,dst k Bk ,θ k ,τ k ,ρ k ), where src k ∈V and dst k ∈V represent the business flow f k The source node and the destination node, Indicates flow f k Required bandwidth and They represent flow f respectively k The deadline and base period, Indicates flow f k Prioritization. Select m paths that meet the needs of m service flows for transmission.

[0023] Furthermore, as an improvement to the efficient routing method for ultra-high bandwidth service flows in time-sensitive networks provided by this invention, the specific process of service modeling in step S3 is as follows:

[0024] S31: After a data flow enters the switch, the switch checks the PCP field and VLAN ID field within the IEEE 802.1Q VLAN tag in the Ethernet frame header. By identifying these fields, the switch determines the attribute information or statistical information of the data flow to further classify and determine its traffic type and priority information.

[0025] S32: Based on the latency attribute information of the data stream, the data stream is first divided into different traffic types, including TT stream, AVB stream and BE stream.

[0026] S33: Based on step S32, data flows in the TT, AVB, and BE streams whose bandwidth reaches a certain threshold of the link bandwidth are additionally classified as ESF, ultimately dividing the data flows into TT, ESF, AVB, and BE streams. Here, let F = {f1, f2, ..., f k} represents the set of TSN service flows, including subsets of TT flows, ESF flows, AVB flows, and BE flows: F = F TT ∪F ESF ∪F AVB ∪F BE .

[0027] S34: After classification, the data streams will enter their respective priority queues in the switch. Each queue represents a traffic type and priority. These queues are used to queue according to priority to ensure that high-priority traffic (TT streams) can be processed and transmitted faster.

[0028] S35: Finally, the CNC establishes its switch queue information table according to the transmission rules of each type of service flow. The switch selects non-conflicting outgoing ports to queue and forward each flow according to the queue information table. This ensures that the transmission of various types of flows will not conflict, so as to maintain the smooth transmission of the flow.

[0029] Furthermore, in step S4, a multi-objective optimization mathematical model considering multiple routing constraints is established. This model aims to minimize the end-to-end delay of the RSF, minimize the available bandwidth rate, and simultaneously find paths with larger remaining bandwidth and fewer flows for the ultra-high bandwidth flow (UBF). The specific objective function is as follows:

[0030] S41: Objective function 1 is to minimize the total end-to-end delay of the TT stream, AVB stream, and BE stream.

[0031] Δ1=ω1·Delay(TT)+ω2·Delay(AVB)+ω3·Delay(BE)

[0032] In the formula, Δ1 is the total end-to-end delay weighted sum, which is between [0, 1], ω1, v2 and ω3 represent the weight values, which are ω1+ω2+ω3=1, and Delay(TT), Delay(AVB) and Delay(BE) represent the normalized average delay of the TT stream, AVB stream and BE stream respectively.

[0033] Furthermore, the normalized average delay of the TT stream, AVB stream, and BE stream in step S41 can be specifically expressed as:

[0034] S411: The normalized average delay of the TT stream is defined as follows:

[0035]

[0036] In the formula, |F TT | represents the number of TT streams, θ TT It is a constant representing the maximum acceptable delay for the TT stream, Delay(f) k ) represents demand f k End-to-end delay;

[0037] S412: The normalized average delay of an AVB stream is defined as follows:

[0038]

[0039] In the formula, |F AVB | Represents the number of AVB streams, θ AVB It is a constant representing the worst-case end-to-end delay of the AVB stream;

[0040] S413: Focus on finite worst-case BE traffic latency without degrading high-priority traffic performance. The normalized average latency of BE flow is defined as follows:

[0041]

[0042] In the formula, |F BE | represents the number of BE flows, θ BE It is a constant representing the acceptable delay threshold for the BE stream. Additionally, |F TT |<|F AVB |<<|F BE |;

[0043] Furthermore, the requirement f in step S411 k Let F represent a set of TSN streams, including TT streams, AVB streams, BE streams, and UBF streams. Therefore, there is a set of TSN streams: F = {f1, f2, ..., f...} k}, where F≡F TT ∪F AVB ∪F BE ∪F UBF Then the demand f k End-to-end delay Delay(f k This can be specifically represented as follows:

[0044]

[0045] In the formula, Indicates link delay, including transmission delay. Propagation delay and queue delay This indicates that the TSN flow is at node v i and v j The direction of transmission between them and Specifically, it can be expressed as:

[0046] (1)

[0047] (2)

[0048] S42: Objective function 2 is to maximize available bandwidth.

[0049]

[0050] Furthermore, some relevant definitions are given in step S42 to facilitate understanding and use. The specific definitions are as follows:

[0051] (1) Effective route: For f kIn other words, an effective route Defined as meeting its deadline The required route. It can be represented as an ordered list of physical links <(src k ,v i ),(v i+1 ,v i+2 ),......,(v j ,dst k )>,f k It can be found from its source address src k to destination address dst k Traverse these physical links.

[0052] (2) Link load: For each link e ij =(v i ,v j ), defines link e ij The load is loaded using Load(e) ij ) indicates that, via link e ij The sum of the magnitudes of all flows. Formally, its load can be defined as:

[0053]

[0054] In the formula, Represents data stream f k The bandwidth size.

[0055] (3) Link bandwidth utilization: For each link e ij =(v i ,v j ), defines link e ij Bandwidth utilization, using U ij Indicates. Link e ij The bandwidth utilization rate is:

[0056]

[0057] In the formula, lb represents the bandwidth of the link, and all links have the same bandwidth.

[0058] (4) Remaining bandwidth of the route: route Use the remaining bandwidth Representation. Defined as The minimum remaining bandwidth of all links in the network can be expressed as:

[0059]

[0060] (5) TSN flow count of the route: route TSN stream count Representation. Defined as The maximum number of TSN flows across all links in the network can be expressed as:

[0061]

[0062] In the formula lt ij Indicates that through link e ij =(v i ,v j The number of TSN streams is obtained.

[0063] (6) Accessibility of each node: using Describes the k-th flow f k Did node v visit? i , can be defined as:

[0064]

[0065] To prevent loops, each node v i All are in the same flow f k Visit once. Therefore, for each node v i ,have:

[0066]

[0067] (7) Bandwidth utilization: routing Bandwidth utilization Representation. Defined as The maximum bandwidth utilization of all links in the network can be expressed as:

[0068]

[0069] S43: Objective function 3 is to minimize the path blocking rate of UBF. Specifically, it involves selecting paths with larger remaining bandwidth and fewer streams to transmit UBF. The specific objective function is as follows:

[0070]

[0071] In the formula, This represents the maximum remaining bandwidth among all valid paths. The conversion value representing the TSN stream number is as follows:

[0072]

[0073] In the formula, This represents the minimum number of TSN flows across all valid paths.

[0074] Furthermore, the objective function established in steps S41, S42, and S43 needs to consider multiple routing constraints to realize a multi-objective routing optimization model. The specific constraints are as follows:

[0075] 1): Delay constraints of TSN data stream

[0076] ①:

[0077] ②:

[0078] ③:

[0079] ④:

[0080] Constraints ① and ② limit the maximum allowed end-to-end delay for TT and AVB streams, respectively; constraint ③ states that BE stream transmission should be completed within the worst-case deadline. Although BE traffic is delay-insensitive, it still needs to be guaranteed to complete transmission within the current time period; constraint ④ guarantees lower end-to-end delay for higher-priority streams. and These are the delays of two different paths.

[0081] 2): Node access degree constraint

[0082] ①:

[0083] ②:

[0084] Among them, constraint ① avoids loops, and a node can only be accessed once by the same data stream. Constraint ② restricts the m data streams to reach the destination node.

[0085] 3): Bandwidth constraints

[0086] ①:

[0087] ②:U ij ≤lb ij ×10%

[0088] ③:

[0089] in, Indicates link e ij The sum of the bandwidth of the streams transmitted on the link. Constraint ① prevents the current link load from not exceeding the link bandwidth; Constraint ② limits the bandwidth utilization of each link to no more than 10% of its capacity to ensure load balancing of the links; Constraint ③ indicates that the ESF bandwidth meets its link bandwidth requirements and does not exceed 10% of the remaining bandwidth.

[0090] 3): Constraints on the number of TSN flows on the path

[0091] ①:

[0092] Among them, constraint ① prevents the number of traffic on the paths allocated by UBF from not exceeding the threshold of Y, thus ensuring the service quality of UBF and avoiding network congestion.

[0093] Furthermore, step S5 employs Transformer-based Deep Reinforcement Learning (TDRL) to obtain a feasible initial solution as an alternative path. Specifically, the TSN Multi-Objective Hybrid Traffic Flow Routing Problem (MOHSFR) is first modeled as a Markov Decision Problem (MDP). Then, the policy network uses a Transformer architecture to model the agent solving the MOHSFR problem. This model mainly includes encoding, decoding, and training processes. By inputting the directed graph of the network, the characteristics and requirements of the flow (known), and the trained Transformer-based DRL model, a set of initial flow paths is output to represent individuals in the solution. These initial solution paths are then used as input for step S6.

[0094] It includes the following specific execution steps:

[0095] S51: Modeling the MOHSFR problem as an MDP:

[0096] S511: Define MDP elements:

[0097] (1) Agent: The DRL model acts as an agent, responsible for the decision path selection action.

[0098] (2) State set s: The state set includes network state S t All possible values ​​of .

[0099] (3) Action set A: The action set includes all possible path selection actions A. t .

[0100] (4) Reward function R: Define a reward function to calculate the reward based on the path performance (such as latency weighting, available bandwidth rate and path blocking degree of UBF).

[0101] S512: MDP modeling. After defining MDP elements, an MDP model is built, including state transition probabilities, reward functions, etc. The state transition function, given the current state and action, ultimately outputs the probabilities of all possible next states. Because the system state changes, the node states, link states, and flow states will change accordingly.

[0102] S52: The encoding process based on the transformer architecture first uses a linear layer to embed the network topology features and each service flow of the TSN into a high-dimensional space and obtain the initial embedding information X. (0) Then X (0) The final information embedding X is obtained by processing through L identical layers. (L) Each layer consists of a multi-head attention layer (MHA), a skip connection layer, a batch normalization (BN) layer, and a fully connected feedforward (FF) layer.

[0103] S53: The decoding process based on the transformer architecture, where the decoder utilizes context information from the encoder to decode at each step, generating a probability vector for selecting the next-hop node for each business flow. Here, the agent selects the path and action at each time step according to the DRL model's strategy, constructs the final path solution, and then outputs the initial solution, returning a set of initial paths generated by the encoder process, which constitute part of the initial population;

[0104] S54: Training process: The policy gradient is used to train the parameters θ in the neural network, which are used to update the parameters of the DRL model. The model provides feedback based on the generated path and performance metrics to optimize the policy network.

[0105] S55: Repeat steps S51-S55 to generate different initial solutions or optimized paths;

[0106] S56: Training ends, outputting a set of initial flow paths to represent individuals in the solution;

[0107] Furthermore, in step S6, the CNC uses the alternative paths selected in step S5 for optimization by NSGA-II to select the preferred path from the transmitter to the receiver that meets various streaming requirements, specifically including the following steps;

[0108] S61: First set the initialization parameters, including population size Size and NSGA-II maximum number of iterations ζ. NSGA-II Crossover rate Cr NSGA-II Variation rate Mr NGGA-II ;

[0109] S62: Transform the initial solution obtained in step S5 into an initial population of NSGA-II to create a parent population P0 of size Size;

[0110] S63: Sort the parent population P0 according to the non-dominant level and assign each chromosome a value equal to its non-dominant level. Set the initial number of cycles to ι = 0.

[0111] S64: Create an offspring population Γ0 on P0 using selection (based on non - domination level), crossover, and mutation operators;

[0112] S65: Determine whether the constraint conditions are satisfied. If the constraint conditions are satisfied, stop and return pι; if not, execute step S66;

[0113] S66: Merge the initial parent population P0 and the offspring population Γ0 to generate a merged population R ι ;

[0114] S67: Identify the non - domination fronts F1, F2, …, F ι from R, and set κ i = 1;

[0115] S68: When |P ι+1 | < Size, calculate the crowding distance of the solutions in F i ; if |P ι+1 | ∪ |F i | ≤ Size, then set P ι+1 = P ι+1 ∪ F i ; otherwise, sort F i according to the crowding distance, and add the least crowded Size - |P i | solutions from F ι+1 to P ι+1 ; then set i = i + 1;

[0116] S69: Determine whether the termination condition is satisfied. If not, set ι = ι + 1, return to step S64, and apply selection (based on non - domination level), crossover, and mutation operators on P ι+1 to create an offspring population Γι + 1; if satisfied, output the solution obtained after NSGA - II optimization. That is, the preferred path that meets various flow requirements from the sender to the receiver is selected.

[0117] ​Compared with existing technologies, this invention has the following significant advantages: (1) This invention specifically optimizes the routing of ultra-high bandwidth flows to ensure that they meet the requirements of high bandwidth and low latency, thereby achieving a more balanced allocation of network resources, reducing the risk of network congestion, and improving overall network performance. (2) This invention can solve the requirements of service load balancing and low latency in TSN local area networks; through intelligent path selection and routing strategies, it distinguishes between ultra-high bandwidth flows and regular service flows to meet their respective QoS requirements. This innovative strategy is very beneficial for applications such as handling large data transmission, immersive video interaction, and generative large model data synchronization. (3) This invention utilizes Transformer-based deep reinforcement learning (TDRL) to generate initial solutions, effectively avoiding the problem of excessively long training time in traditional DRL. The TDRL model generates quickly, providing high-quality initial solutions for the NSGA-II algorithm, improving overall optimization efficiency and success rate. This multi-objective optimization method can comprehensively consider different aspects of network performance, enabling the network to perform well under various traffic requirements. Attached Figure Description

[0118] Figure 1 This is a schematic diagram illustrating the communication forwarding initiated by the service flow in the TSN network architecture of the present invention.

[0119] Figure 2 This is a schematic diagram of a CNC structure for identifying and classifying TSN mixed service flows according to the present invention.

[0120] Figure 3 This is a flowchart illustrating an efficient routing method for multi-objective optimization in time-sensitive networks for mixed-stream transmission, as proposed by the present invention.

[0121] Figure 4 This is a schematic diagram of the first stage of the process of solving feasible initial solutions using transformer-based DRL and optimizing them for NSGA-II.

[0122] Figure 5 This is a schematic diagram of the process of using the obtained feasible initial solution for optimization in NSGA-II according to the present invention. Detailed Implementation

[0123] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0124] Example 1: The communication forwarding process initiated by the service flow in the TSN network structure.

[0125] Due to the high resource consumption of ultra-high bandwidth (UBF) streams, this new scenario places significant pressure on reliable communication. Furthermore, the limited switching capacity and buffer space of current networks easily lead to network congestion and packet loss. To address these issues, this invention proposes an efficient routing method for UUBF streams in Time-Sensitive Networks (TSNs). The core idea for meeting the transmission requirements of mixed services in TSNs is: first, to distinguish UBFs from other data streams to meet their respective needs. Then, a novel routing algorithm is used to select the optimal path for transmission for each service stream. The communication forwarding process of a service stream in a TSN is as follows: Figure 1 As shown.

[0126] The workflow of a service flow from the source node to the destination node via relay nodes is as follows:

[0127] Step 1: Before the network starts running, the Central User Configuration (CUC) initiates a request to retrieve the network physical topology to the Central Network Controller (CNC) through the User Network Interface (UNI).

[0128] Step 2: CNC discovers the TSN network topology, abstracts the TSN network topology into a directed network graph, and returns the result to CUC;

[0129] Step 3: The terminal device sends a TSN connection request and requirement to the CUC through the user configuration protocol, such as which terminal devices need to communicate, the period, size and latency of the TSN stream, etc. The CUC sends the connection request and requirement to the CNC through UNI.

[0130] Step 4: The controller (CNC) divides these service flows into TT flows, AVB flows, and BE flows based on the QoS latency requirements of different flows; at the same time, when the bandwidth of the data flow reaches a certain threshold of the link bandwidth, it will be classified as UBF and mapped to different priority queues of the switch.

[0131] Step 5: Based on the network physical topology and network requirements, the CNC uses the multi-objective optimization and efficient routing method for hybrid flow transmission proposed in this scheme to perform route calculations and obtain relevant path information that meets the connection requirements and is feasible. The route calculation is based on T-DRL to obtain a feasible initial solution, which is then used for optimization in NSGA-II.

[0132] Step 6: If such a path and transfer table exist, the CNC configures all TSN switches along the calculated path using the calculated transfer table;

[0133] Step 7: The CNC sends the transfer table to the CUC, and the CUC forwards it to the sending device;

[0134] Step 8: The sending device starts sending the TSN stream to the receiving device along the selected path according to the calculated transmission table.

[0135] Example 2: The process of CNC identifying and classifying TSN mixed service flows

[0136] To make step four in Example 1 clearer, this example uses a schematic diagram of the process of controlling the CNC to identify and classify TSN mixed service flows as an example. Figure 2 As shown, this paper introduces the identification and classification process of TSN mixed service flows to meet the transmission requirements of various service flows. The specific identification and classification process is as follows:

[0137] Step 1: The CNC receives the TSN service flow connection request sent by the CUC through the UNI, identifies the QoS attribute requirements of the service flow, and then extracts the characteristic requirements of the service flow, including latency and bandwidth.

[0138] Step 2: CNC defines the classification rules for service flows and assigns PCP and VID identifiers to each service flow. The specific classification rules are as follows: first, extract the latency attribute of the service flow for initial classification; then, based on the service flows classified in the first classification, extract the bandwidth attribute for further classification.

[0139] Step 3: Based on the classification rules, the classifier performs specific classification operations on the service flows. First, based on the information extracted from the features, it classifies them into TT streams, AVB streams, and BE streams according to the latency attribute; then, based on this, it further classifies them into TT streams, UBF streams, AVB streams, and BE streams according to the bandwidth attribute.

[0140] Step 4: Finally, the controller maps the classified service flows to different priority queues of the switch according to the assigned PCP and VID, waiting for further processing, and selects non-conflicting switch outgoing ports to forward the flows.

[0141] Step 5: Service flows enter their respective priority queues for queuing, waiting for subsequent processing to select a non-conflicting switch outgoing port for forwarding.

[0142] Example 3: Detailed process of a multi-objective optimization and efficient routing method for hybrid stream transmission in time-sensitive networks.

[0143] To make step five in Example 1 clearer, this example uses a flowchart of a multi-objective optimization method for efficient routing of hybrid streams in time-sensitive networks as an example. Figure 3 As shown, this paper introduces an efficient routing algorithm for TSN mixed service flows to meet the transmission requirements of various service flows. The proposed algorithm mainly consists of three stages:

[0144] The first stage is the network modeling stage, where the controller (CNC) performs network modeling based on the network topology and service flow attributes, which serves as input for the subsequent two stages.

[0145] The second stage is a deep reinforcement learning stage based on transformers. When the CNC receives a TSN service flow connection request sent by the CUC through UNI, the CNC first executes the T-DRL algorithm to select the optimal path that meets the requirements of various service flows based on the network topology and service flow requirements. Then, the solution generated by T-DRL is used as the initial solution for the optimization stage of the NSGA-II algorithm to improve the quality of the solution and obtain an ideal Pareto line. At the same time, this method can also solve the problem of the time consumption of the initial solution when the traditional heuristic algorithm encounters complex constraint problems.

[0146] The third stage is the NSGA-II algorithm optimization stage. The CNC encodes the solution generated by the T-DRL in the second stage into the NSGA-II chromosome form, creating an initial population P0. Then, the parent population P0 is sorted according to the non-dominated level, and each chromosome is assigned a value equal to its non-dominated level, with the initial loop count set to ι = 0. Selection (based on non-dominated level), crossover, and mutation operators are used to create a descendant population Γ0 on P0. The system checks if the constraints are met; if so, it stops and returns to P1; otherwise, the initial parent population P0 and the descendant population Γ0 are merged to generate a merged population R. ι Then, a new population P is created through selection, crossover, and mutation operations. ι The process continues until the maximum number of iterations is met, obtaining an ideal Pareto optimal solution; finally, the transmission table results calculated for each service flow are sent to the TSN switch.

[0147] The specific algorithm flow of the efficient routing method for ultra-high bandwidth service flows in time-sensitive networks is as follows:

[0148] Step 1: The CNC discovers the TSN physical network topology based on LLDP, models it as a directed graph G=(V,E) using network modeling methods, and returns the result to the CUC;

[0149] Step 2: The terminal system configures the communication terminal (src) through the user configuration protocol. k ,dst k ), base period τ of business flow k Deadline θ k Bandwidth size B k Priority ρ k Population size (Size), NSGA-II maximum number of generations (ζ) NSGA-II Maximum number of iterations ζ in T-DRL T-DRLThe system sends information to the CUC to obtain all information in the network, and then the CUC passes these communication requests to the CNC through the UNI.

[0150] Step 3: When the CNC receives a service request, it first generates a feasible solution based on the T-DRL algorithm according to the network topology and the requirements of various flows, that is, selects a path that meets the requirements for transmission for each service flow;

[0151] Step 4: The CNC uses the solution generated in Step 3 based on the T-DRL algorithm as the initial solution for NSGA-II, and then performs subsequent optimization through selection, crossover and mutation operations to obtain the ideal Pareto optimal solution;

[0152] Step 5: CNC checks if there are any service flows for which the transmission path from the source to the destination has not yet been calculated. If so, return and execute steps 3 to 6, and save the preferred path for each service flow to the transmission path table. If not, continue to execute step 6.

[0153] Step 6: The controller CNC configures the calculated service flow transfer table to the switch via the NETCONF protocol, and at the same time sends its transfer table to the terminal device through the CNC.

[0154] Example 4: Solving feasible initial solutions using transformer-based DRL and applying it to further optimize the NSG-II algorithm.

[0155] To make the second stage in Example 3 clearer, let's take the flowchart of the first stage, which uses the transformer-based DRL to solve for a feasible initial solution and then uses it for subsequent optimization in the NSG-II algorithm, as an example. Figure 4 As shown. The steps to solve for the initial solution first include modeling, encoding, decoding, and training. A detailed description follows:

[0156] Process 1: When solving for MOHSFR using deep reinforcement learning, we model it as an MDP, where the state S, action A, state transition function, and reward R are defined as follows:

[0157] 1): State Space

[0158] The routing process involves multiple system states. The system state in step t consists of the service flow state, the SW node state, and the link state, as described below. The service flow state is determined by its source node src. k , destination node dst k Current position node loc k Required bandwidth B k Cumulative delay γ k and priority ρ kComposition, represented as The node state consists of the remaining bandwidth of all links connected to the SW node, and can be represented as follows: The node state includes the node's M priority queues, which can be represented as follows: Each element represents a node v. i The length of the Mth priority queue. Therefore, the system state space can be represented as

[0159]

[0160] 2): Space for Action

[0161] In the MOHSFR problem, the agent must allocate a path for each traffic flow from the source node to the destination node. First, the agent selects the next hop for each traffic flow in the network from its neighboring nodes. Then, the traffic flow is transmitted to the next node and enters the appropriate priority queue. Therefore, the state... Indicates flow f k A node will be reached in step t, where It is a vector representing the flow f from the business process. k The next hop selected from the neighboring nodes, and Represents node v i The flow f k Choose as the next jump. Represents node v i Not selected. The sequence of actions from step 0 to the last step T can be represented as {a 0 ,a 1 ,…,a T}, Where n represents the number of service flows (SWs). Here, each service flow f... k Initially located in the source node src k , using a 0 =0 indicates that the destination node dst is reached eventually. k , using a T =-1 indicates. a t ∈{1,…,n} represents the business flow f k Flow to SW node.

[0162] 3): Reward function

[0163] By executing action {a 0 ,a 1 ,…,a T This allows us to obtain the routing paths of all data streams, as well as their arrival times and link bandwidth usage. The total reward function is calculated based on the objective function established in steps S41, S42, and S43, specifically represented by the following formula:

[0164] Reward=-{ω6Δ(1)+ω7Δ(2)+ω8Δ(3)}.

[0165] Where ω6+ω7+ω8=1, Δ(1), Δ(2) and Δ(3) are three objective functions.

[0166] 4): State transition function

[0167] Current system state S t Based on the currently executing action a t Update to the next state S t+1 Once the service flow selects the next hop, the link state will also be updated as the action is executed, which can be represented as:

[0168]

[0169] In addition, the transmission latency of TSN service flows and the number of flows on their paths will also increase, as shown below:

[0170]

[0171] Process 2: Information encoding process based on transformer architecture:

[0172] 1) The encoder consists of one linear layer and L identical layers. First, the encoder uses the linear layer to embed the network topology features and each traffic flow of the TSN into a high-dimensional space and obtain initial embedding information. Furthermore, we employ separate linear projections to handle different information regarding TSN traffic flows and network topology, as shown below:

[0173]

[0174] The network topology requires embedding node information, link bandwidth capacity, and maximum load information, while TSN service flows require embedding source address, destination address, priority, deadline, and bandwidth requirement information. The operation [·,·,…,] represents concatenating information from different dimensions together, and... These are trainable linear projection parameters, where d x X represents (0) The dimension. In the algorithm of this invention, d x The value is set to 128.

[0175] 2): Then, the initial information is embedded in X. (0) After processing through L identical layers, the final information embedding X is obtained. (L)Each layer consists of a multi-head attention layer (MHA), a skip connection (SC) layer, a batch normalization (BN) layer, and a fully connected feedforward (FF) layer. The input to layer l is the output of the previous layer, denoted as... X (l-1) First, the attention value is calculated using MHA, which employs a dimension of d. k The H header, and let H = 8, For each head h∈{1,2,…,H}, the first step is to query... key Sum The calculation is as follows:

[0176]

[0177] in These are the trainable parameters of the l-sublayer.

[0178] 3): Obtain the Head Attention Value using the softmax function. Then Connect the projection to dimension d x Obtain the multi-head attention value MHA(h) (l-1) The main formulas are as follows:

[0179]

[0180] in These are the trainable parameters of sublayer l.

[0181] 4): Further processing of the attention value using SC and BN yields... Finally, the output X, which is embedded in the l-layer information, is obtained after passing through the FF layer and SC layer. (l) The calculation process is shown in the following formula:

[0182]

[0183] Process 3: Information Decoding Process Based on Transformer Architecture

[0184] During the decoding process, the decoder outputs a probability vector for each node to be selected. Various constraints are established here, indicating that the existence of these constraints may prevent certain incoming links or nodes from being accessed. Therefore, a blocking rule is used to exclude nodes or incoming links that cannot be accessed in the current step due to the constraints. The main blocking rules include: (1) switch nodes accessed by the same data flow are blocked; (2) nodes whose data flow bandwidth exceeds the remaining bandwidth of the link are blocked; (3) nodes whose data flow arrival delay exceeds the deadline are blocked. Whether the i-th fixed point is blocked in step t is defined as follows:

[0185]

[0186] Furthermore, the main decoding process of process 3 includes:

[0187] 31): The decoder uses context information from the encoder to decode at each step, generating a probability vector for selecting the next-hop node for each traffic flow. At each step t, the agent determines the next-hop node based on the current state s. t Make a decision t The context obtained by the decoder Including the entire graph X (L) The embedded information, and the embedded information of the business flow at the previous node position. Specifically, it is expressed as follows:

[0188] 32): Subsequently, the current state s is calculated through the MHA layer. t Context information below Wherein, the query vector of MHA is The context information is used to define the key vector and value vector as the node's embedding.

[0189]

[0190] 33): Calculate the probability vector of each node being selected through an attention mechanism, i.e., calculate st g Compatibility with all nodes:

[0191]

[0192] 34): Finally, combining the data flow and the compatibility vector between nodes, the softmax function outputs the probability vector for each hop node selection:

[0193] p t =softmax(ξ) t +C·Mask t )

[0194] Here, C is a very large negative number. The decoder will repeat this process at each step, selecting the next hop node, until the requirements of all streams are met and they have all reached their destination node.

[0195] Step 4: The training process applies the baseline policy gradient method to the neural network to train parameters θ. This algorithm consists of a policy network and a baseline network, both with the same network structure. The only difference is that the policy network selects actions based on sampling the probability vector, while the baseline network selects the action with the highest probability (greedy policy). The gradient of the loss function can be defined as: Where R(π)θ ) and R(π BL The values ​​() represent the rewards obtained by the policy network and the baseline network, respectively. During each training iteration, the model undergoes a paired t-test. If the test result is significant at 95% confidence, the parameters of the policy network are replaced with those of the baseline network. This step ensures that the updated parameters achieve a statistically significant improvement over the previous parameters, thereby further enhancing model performance.

[0196] The solution generated based on TDRL is used as the initial solution for the NSGA-II algorithm for further optimization.

[0197] To make the third stage in Example 3 clearer, a flowchart illustrating the process of further optimizing the solution generated based on TDRL as the initial solution of the NSGA-II algorithm is used as an example. Figure 5 As shown. The specific steps of the algorithm are as follows:

[0198] Step 1: Initialization of the NSGA-II population

[0199] The T-DRL algorithm yielded an initial set of feasible routes and path selection solutions. Each solution represents a route path for traffic flows in the TSN to satisfy a given optimization objective. The initial solutions generated by DRL were then transformed into a population representation adapted to NSGA-II, that is, the route paths were encoded as chromosomes. Typically, an NSGA-II population is a set of individuals, each representing a solution. Thus, a single chromosome represents the complete solution to the problem and can be defined as follows:

[0200] x = [path1,path2,...,path] m ]

[0201] Where path1, path2, path3 and path m This represents the path allocation for different flows. Simply put, genes in a chromosome can be viewed as a set of routing paths, with each gene representing a data flow and encoding its path information. The combinations of these path choices constitute the chromosome, thus determining the data flow routing configuration throughout the network. Gene encoding uses integer encoding, where each integer represents a node on the path, as shown below:

[0202] path i =[node1,node2,...,node m ]

[0203] Through the above process, a set of solutions generated by T-DRL can be integrated into the initial population of NSGA-II, creating a parent population of size Gen: P0 = {x1, x2, ..., x...} Gen}

[0204] Step 2: Calculate the fitness value for different targets on each chromosome, compare each chromosome with other chromosomes to check if it is dominant, and sort the parent population P0 according to the non-dominant level of the chromosomes, assigning each chromosome a fitness value equal to its non-dominant level. Set the initial number of iterations to ι = 0; specifically:

[0205] First, a fitness function algorithm needs to be defined: the quality of a chromosome is generally evaluated using its fitness value. In this invention, we focus on minimizing the end-to-end latency and maximum link bandwidth utilization of the GSF in the TSN network, as well as minimizing the path degree of the ESF. Therefore, a separate fitness function value is defined for each optimization objective. The quality of each solution is evaluated according to different objectives, and these values ​​are set to be the same as the corresponding objective function. Therefore, the fitness function is defined as follows:

[0206] Fitness1 = Δ1;

[0207] Fitness2 = Δ2;

[0208] Fitness3 = Δ3;

[0209] Then, a fast non-dominated sorting rule is defined: after generating the initial population and evaluating multiple fitness values ​​of chromosomes, a non-dominated sorting algorithm is applied to the population to identify all non-dominated fronts {Z1, Z2, Z3, ...}, that is, all non-dominated fronts are classified according to the dominance level of chromosomes. In this invention, the dominance of two chromosomes CH(i) and CH(j) is first identified by comparing their end-to-end delay and path degree. Chromosome CH(i) dominates CH(j) if the following conditions are met.

[0210] (1): The objective functions Δ1(i) and Δ3(i) are better than Δ1(j) and Δ3(j) respectively, regardless of the value of Δ1. In other words, (Δ1(i)>Δ1(j))∧(Δ3(i)>Δ3(j));

[0211] (2): The objective functions Δ1(i) and Δ3(i) reach their minimum values, while at least one of Δ1(j) and Δ3(j) does not reach its minimum value, regardless of the value of Δ2. In other words, (Δ1(i) = min1 ∧ Δ3(i) = min3) ∧ (Δ1(j) > min1 ∨ Δ3(j) > min3), (where min1 and min3 are the minimum values ​​of the objective functions Δ1 and Δ3, respectively). Otherwise...

[0212] (3): CH(i) is superior to or equal to CH(j) on all objectives (Δ1, Δ2, Δ3) and strictly dominates CH(j) on at least one objective in (Δ1, Δ2, Δ3).

[0213] Step 3: Use selection (based on non-dominated level), crossover, and mutation operators to create a progeny population Γ0 of size Size on P0; specifically:

[0214] (1) Selection Operator: The main idea of ​​the selection operator is to prioritize the most fit individuals in the population during the crossover process. In the selection phase, a pair of chromosomes is selected from the population for crossover and mutation operations. In traditional single-objective problems, selection typically depends on fitness values, meaning chromosomes with higher fitness values ​​are more likely to be selected. However, for the proposed algorithm, since the problem is multi-objective, the selection process adopts a method based on dominance level and crowding distance, i.e., selecting chromosomes with higher crowding distances and lower non-dominated levels. Furthermore, a sufficient number of chromosomes must be selected during the selection phase so that they can generate a number of child chromosomes (Size) after the crossover operation.

[0215] (2) Crossover Operator: The crossover operator combines the genes of two individuals by extracting information from both parents and recombineing it into new offspring, allowing the inheritance of this information. The crossover operator proposed in this invention adopts a single-point crossover method, allowing non-repeating routes of different individuals at the same location to be exchanged, thus avoiding the repetition of the same route. The crossover rate is set to 90% here.

[0216] (3) Mutation Operator: The mutation operator allows the population to acquire new genes to maintain population diversity and prevent the algorithm from converging prematurely. This algorithm randomly selects two positions within the chromosome from the corresponding route set and exchanges the information at these positions. The mutation rate is set to 10% here.

[0217] Step 4: Determine if the constraints are met. If the constraints are met, stop and return to P1; if the constraints are not met, proceed to Step 5.

[0218] Step 5: Merge the initial parent population P0 and the offspring population Γ0 to generate a merged population R. ι ;

[0219] Step 6: Then apply the following to population R ι Fast nondominated sorting is applied to classify the nondominated fronts {Z1, Z2, Z3, ...} at different nondominated levels. Then, a next-generation population P of size Size S is created based on the crowding distance of the sorting. ι+1 The specific steps are as follows:

[0220] The first step is to start from R ιIdentify the non-dominated frontiers F1, F2, ..., F κ and set i = 1;

[0221] The second step is when |P ι+1 When | < Size, calculate F i The congestion distance in the middle;

[0222] Third step, if |P ι+1 |∪|F i |≤Size, then let P ι+1 =P ι+1 ∪F i ;

[0223] Step 4, otherwise (|P ι+1 |∪|F i |>Size), based on the crowding distance for F i Sort and add from F i To P ι+1 Least crowded Size-|P ι+1 |One solution;

[0224] Fifth step, then set i = i + 1; return to steps two through five and continue until a new population P is formed. ι+1 The process ends when the population size meets the required size (Size).

[0225] Step 6: Generate a new population P with a population size of Size. ι+1 And let ι = ι + 1;

[0226] Step 7: Determine if the termination condition ι > ζ is met. NSGA-II If the condition is not met, return to step 3, i.e., at P. ι+1 The above uses selection (based on non-dominated level), crossover, and mutation operators to create the offspring population Γ. ι+1 If satisfied, the output is the Parto optimal solution obtained after NSGA-II optimization, which selects the preferred path from the sender to the receiver that meets various streaming requirements.

[0227] This invention, while meeting service flow requirements and constraints, minimizes the path congestion rate of UBF, minimizes the average end-to-end latency of RSF, and maximizes the available bandwidth, thereby obtaining a set of Pareto optimal solutions. Compared with traditional methods, the method of this invention can significantly improve network performance in scenarios where UBF and RSF are deployed in combination, meeting the high requirements of time-sensitive networks.

[0228] In summary, this invention significantly improves solution quality and convergence performance through a multi-objective optimization method, solving the problem of long initial solution processing time in traditional heuristic algorithms. When UBF (Unified Burst Frequency) exists, the method of this invention can achieve better load balancing, reduce network congestion and path blocking, and improve the overall network stability and reliability. It effectively reduces the average end-to-end latency of RSF (Relational Burst Frequency), meeting the latency requirements of different types of service flows and improving user experience. By maximizing available bandwidth, it improves the utilization efficiency of network resources, avoids bandwidth waste, and ensures that the high bandwidth requirements of UBF are fully met. The method of this invention can be flexibly applied to ultra-high bandwidth service flow routing optimization in various scenarios such as cloud computing, communication, and network services, and has broad application prospects.

[0229] It should be noted that the above-described embodiments are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be pointed out that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. An efficient routing method for ultra-high bandwidth service flows in time-sensitive networks, characterized in that, include: S1: The centralized user configuration CUC initiates a request to retrieve the network physical topology to the network central controller CNC. The CNC discovers the Time Sensitive Network (TSN) topology and abstracts it into a directed network graph, returning it to the CUC. S2: The terminal system sends TSN connection requests and resource requirements to the CUC through the user configuration protocol, and the CUC passes these connection requests and resource requirements to the CNC; S3: The CNC performs service modeling at the switch based on service flow attributes, assigns PCP and VLANID fields to each service flow, classifies service flows according to service characteristics, and maps them to different priority queues of the switch. The service flows are TSN flows. The specific classification rules are as follows: First, the latency attribute of the service flow is extracted and classified for the first time, dividing the service flow into different traffic types, including time-triggered (TT) flow, audio / video bridging (AVB) flow, and best-effort (BE) flow. Then, based on the service flows after the first classification, the bandwidth attribute is extracted and classified again. Service flows in TT flow, AVB flow, and BE flow whose bandwidth reaches a certain threshold of the link bandwidth are additionally classified as ultra-high bandwidth (UBF) flow. Finally, the service flows are divided into TT flow, UBF flow, AVB flow, and BE flow. S4: Obtain basic information about the entire network, including the TSN network topology map and service flow information set, and establish a multi-objective optimization mathematical model that considers multiple routing constraints based on this basic information; Its constraints include: 1) Delay constraints for service flows, including delay deadline constraints for TT flow, AVB flow, BE flow, and high-priority service flows; 2) Node access constraint: a node can only be accessed once by the same business flow, and all business flows must reach the destination node. 3) Bandwidth constraints: The current link load is limited to the link's bandwidth; the bandwidth utilization of each link does not exceed a certain ratio of its capacity; the UBF flow bandwidth meets its link bandwidth requirements and does not exceed a certain ratio of the remaining bandwidth of the effective route; a valid route... Defined as meeting the deadline Required route; remaining bandwidth of the valid route is The minimum remaining bandwidth of all links in the network; 4) Constraints on the number of service flows on the path: the number of service flows on a valid route cannot exceed the threshold. The multi-objective optimization mathematical model considering multiple routing constraints is established, and the specific objective function is as follows: 1) Objective function 1 is to minimize the total end-to-end delay of the TT stream, AVB stream and BE stream, and is the weighted average of the TT stream, AVB stream and BE stream; 2) Objective function 2 is to maximize the available bandwidth. 3) Objective function 3 is to minimize the path blocking rate of UBF, by selecting paths with larger remaining bandwidth and fewer traffic flows to transmit UBF; S5: CNC performs routing calculations based on the network physical topology, network requirements, and a multi-objective optimization mathematical model containing multiple routing constraints to obtain relevant path information that meets connectivity requirements and is feasible. The routing calculation first uses Transformer-based deep reinforcement learning T-DRL to obtain a feasible initial solution and use it as an alternative path. S6: The CNC selects the preferred path from the candidate paths selected by the routing calculation to the receiving end that meets the service flow requirements. S7: CNC traverses to see if there are any uncalculated paths for the sending and receiving ends. If so, return to steps S3-S6; otherwise, proceed to step S8. S8: The CNC encapsulates the calculated path results into a transfer table configuration for the TSN switch and sends it to the TSN terminal device via the CNC.

2. The efficient routing method for ultra-high bandwidth service flows in time-sensitive networks according to claim 1, characterized in that, Step S1 includes: Before the network starts operating, the CUC initiates a request to the Network Central Controller (CNC) through the User Network Interface (UNI) to retrieve the TSN network topology information. The CNC then discovers the TSN network topology and node information using the Link Layer Discovery Protocol (LLDP), models it as a directed TSN graph G = (V, E), where V and E represent the node set and edge set, respectively, and returns the result to the CUC. The specific modeling process is as follows: The TSN network topology is modeled as a directed graph G = (V, E), where V ≡ (v1, v2, ..., v...). n+λ () represents a set of n switch (SW) nodes and a set of λ terminal device (ES) nodes, where E = {e ij |i,j∈n,i≠j} is the set of physical links, where e ij Represents node v i and node v j The links formed by the connections; TSN supports full-duplex, and each full-duplex link is treated as two separate directed links. ij =(v i ,v j ) and e ji =(v j ,v i ); A∈R n×n Let e ​​represent an adjacency matrix consisting of elements 0 and 1, where if e ij ∈E, then A ij =1, if Then A ij =0; each link e ij Defined by triples <lb ij ,ld ij ,lt ij >, among which It is link e ij bandwidth capacity; The link delay, expressed in μsec, is represented by v. i transmission delay pd ij Queuing delay qd ij Link (v) i ,v j The propagation delay td ij composition; To pass through link (v i ,v j The number of business flows.

3. The efficient routing method for ultra-high bandwidth service flows in time-sensitive networks according to claim 2, characterized in that, Step S2 includes: the terminal system configuring the communication terminal (src) via a user configuration protocol. k ,dst k ), base period τ of business flow k Deadline θ k Bandwidth size B k Priority ρ k Population size (Size) and maximum number of iterations (ζ) in NSGA-II NSGA-II Maximum number of iterations ζ in T-DRL T-DRL This information is sent to the CUC, which then forwards these connection requests and resource requirements to the CNC via the UNI, as follows: Each communication task is abstracted as a service flow, and these service flows are defined by tuples f. k ≡(src k ,dst k ,Β k ,θ k ,τ k ,ρ k ), where src k ∈V and dst k ∈V represent the business flow f k The source node and the destination node, Indicates flow f k Required bandwidth and These represent the business flow f respectively. k The deadline and base period, Represents the business flow f k Prioritize; select m paths that meet the needs of m service flows for transmission.

4. The efficient routing method for ultra-high bandwidth service flows in time-sensitive networks according to claim 3, characterized in that, The specific process of business modeling in step S3 is as follows: S31: After a service flow enters the switch, the switch checks the PCP field and VLAN ID field in the IEEE 802.1Q VLAN tag in the Ethernet frame header; by identifying these fields, it determines the attribute information or statistical information of the service flow, so as to further classify and determine its traffic type and priority information; S32: Based on the latency attribute information of the service flow, the service flow is divided into different traffic types, including TT flow, AVB flow and BE flow; S33: Based on step S32, service flows whose bandwidth reaches a certain threshold of the link bandwidth in TT, AVB, and BE flows are additionally classified as UBF flows, ultimately dividing the service flows into TT, UBF, AVB, and BE flows; let F = {f1, f2, ..., f k } represents the set of business flows, where k is an integer, and F includes subsets T and the flow set F. TT UBF stream set F UBF AVB stream set F AVB and BE stream set F BE F = F TT ∪F UBF ∪F AVB ∪F BE ; S34: After classification, the service flows enter their respective priority queues in the switch. Each queue represents a traffic type and priority, and these queues are used to queue according to priority. S35: The CNC establishes its switch queue information table according to the transmission rules of each type of service flow. The switch selects non-conflicting outgoing ports to forward each service flow according to the queue information table.

5. The efficient routing method for ultra-high bandwidth service flows in time-sensitive networks according to claim 4, characterized in that, The objective function in step S4 is as follows: S41: Objective function 1 is to minimize the total end-to-end delay of the TT stream, AVB stream, and BE stream, where Δ1=ω1·Delay(TT)+ω2·Delay(AVB)+ω3·Delay(BE), In the formula, Δ1 is the total end-to-end delay weighted sum, which is between [0, 1], ω1, ω2 and ω3 represent the weight values, ω1+ω2+ω3=1, and Delay(TT), Delay(AVB) and Delay(BE) represent the normalized average delay of the TT stream, AVB stream and BE stream, respectively. S42: Objective function 2 is to maximize available bandwidth. Where U is the bandwidth utilization rate, defined as The maximum bandwidth utilization of all links in the network is expressed as: A valid route Defined as meeting the deadline The required route is represented as an ordered list of physical links <(src k ,v i ),(v i+1 ,v i+2 ),......,(v j ,dst k Business flow f k From the source node src k ∈V to the destination node dst k All nodes ∈V are included in this list; the bandwidth utilization and link load of the link are defined as follows: (1) Link bandwidth utilization: For each link e ij =(v i ,v j ), use U ij Indicates link e ij Bandwidth utilization, link e ij The bandwidth utilization rate is: In the formula, lb represents the bandwidth of the link, and e represents the bandwidth of each link. ij The bandwidth is the same; (2) Link load: For each link e ij =(v i ,v j ), use Load(e ij ) indicates link e ij The load, i.e., the load passing through link e ij The sum of the sizes of all business flows is defined as: In the formula, Represents the business flow f k The bandwidth size; Represents the business flow f k At node v i and v j The direction of transmission between them S43: Objective function 3 is to minimize the path blocking rate of UBF. This is achieved by selecting paths with larger remaining bandwidth and fewer traffic flows to transmit UBF. The specific objective function is as follows: In the formula, ω3 and ω4 represent weights, and ω3 + ω4 = 1; This represents the maximum remaining bandwidth among all valid routes, where the remaining bandwidth of a valid route is defined as... The minimum remaining bandwidth of all links in the network is expressed as: and The conversion value representing the number of business flows is as follows: In the formula, This represents the minimum number of service flows among all valid routes. express The maximum number of service flows on all links in the middle, lt ij Indicates via link e ij =(v i ,v j The number of business flows is obtained. .

6. The efficient routing method for ultra-high bandwidth service flows in time-sensitive networks according to claim 5, characterized in that, The normalized average delay of the TT stream, AVB stream, and BE stream in step S41 is specifically expressed as follows: S411: The normalized average delay of the TT stream is defined as follows: In the formula, |F TT | represents the number of TT streams, θ TT Delay(f) is a constant representing the maximum acceptable delay for the TT stream. k ) represents demand f k End-to-end delay; S412: The normalized average delay of an AVB stream is defined as follows: In the formula, |F AVB | Indicates the number of AVB streams; θ AVB It is a constant representing the worst-case end-to-end delay of the AVB stream; S413: Focus on finite worst-case BE traffic latency without degrading high-priority traffic performance; the normalized average latency of BE flow is defined as follows: In the formula, |F BE | Indicates the number of BE streams; θ BE It is a constant representing the acceptable delay threshold for the BE stream, where |F TT |<|F AVB |<<|F BE |; In step S411, the requirement f k End-to-end delay Delay(f k The following is represented: In the formula, Indicates link delay, including transmission delay. Propagation delay and queue delay This indicates that the TSN flow is at node v i and v j The direction of transmission between them and Specifically, it can be expressed as: (1) (2) 7. The efficient routing method for ultra-high bandwidth service flows in time-sensitive networks according to claim 6, characterized in that, The objective function established in steps S41, S42, and S43 needs to consider multiple routing constraints to realize a multi-objective routing optimization model. The specific constraints are as follows: 1) Delay constraints of business flow ①： ②： ③： ④： Among them, constraints ① and ② limit the maximum allowed end-to-end delay for TT and AVB streams, respectively; constraint ③ requires that the transmission of BE streams should be completed within the worst-case deadline, ensuring its completion within the current time period; constraint ④ guarantees that streams with higher priority have lower end-to-end delays. and These are the delays of two different paths; 2): Node access degree constraint ①： ②： in, f represents the k-th business flow k Did node v visit? i Defined as: Each node v i All are in the same business flow f k Visit once; for each node v i ,have: f represents the k-th business flow k Did you visit your destination node Dst? k Here, constraint ① avoids loops, meaning a node can only be accessed once by the same business flow, and constraint ② restricts all m business flows to reach the destination node. 3): Bandwidth constraints ①： ②:Ui j ≤lbi j ×10%, ③： in, Indicates link e ij The sum of the bandwidths of the streams transmitted on the network. This indicates that the business flow is at node v i and v j The transmission direction between them, U ij Indicates link e ij Bandwidth utilization; Represents the business flow f k The bandwidth size; constraint ① limits the current link load to no more than the link bandwidth; constraint ② limits the bandwidth utilization of each link to no more than 10% of its capacity; constraint ③ indicates that the UBF flow bandwidth meets its link bandwidth requirements and does not exceed 10% of the remaining bandwidth of the effective route. 4): Constraints on the number of business flows on the path ①： Among them, constraint ① restricts the number of traffic on the path allocated by UBF to not exceed the threshold γ.

8. The efficient routing method for ultra-high bandwidth service flows in time-sensitive networks according to claim 7, characterized in that, Step S5 describes using Transformer-based improved deep reinforcement learning (T-DRL) to obtain a feasible initial solution as an alternative path, specifically as follows: The multi-objective hybrid traffic flow routing problem (MOHSFR) of TSN is modeled as a Markov decision problem (MDP). Then, the policy network adopts a Transformer architecture to model the agent solving MOHSFR. This model includes an encoding process, a decoding process, and a training process. By inputting the directed graph of the network, the characteristics of the flow, the known requirements, and the trained Transformer-based DRL model, a set of initial flow paths is output to represent the individuals in the solution. Then, the initial solution paths are used as the input for step S6. The specific execution steps include: S51: Modeling the MOHSFR problem as an MDP: S511: Define MDP elements: (1) Agent: The DRL model acts as an agent, responsible for the decision path selection action; (2) State set S: The state set includes network state S t All possible values; (3) Action set A: The action set includes all possible path selection actions A. t ; (4) Reward function R: Define the reward function to calculate the reward based on the path's performance; S512: MDP modeling. After defining MDP elements, an MDP model is built, including state transition probabilities and reward functions. The state transition function outputs the probabilities of all possible next states based on the given current state and action. S52: The encoding process based on the transformer architecture first uses a linear layer to embed the network topology features and each service flow of the TSN into a high-dimensional space and obtain the initial embedding information X. (0) Then X (0) The final information embedding X is obtained by processing through L identical layers. (L) Each layer consists of a multi-head attention layer (MHA), a skip connection layer, a batch normalization (BN) layer, and a fully connected feedforward (FF) layer. S53: The decoding process based on the transformer architecture. The decoder uses context information from the encoder to decode at each step and generates a probability vector for selecting the next hop node for each business flow. The agent selects the path and action at each time step according to the strategy of the DRL model, constructs the final path solution, and then outputs the initial solution. It returns a set of initial paths generated by the encoder process, which constitute part of the initial population. S54: Training process: The policy gradient is used to train the parameters θ in the neural network, which are used to update the parameters of the DRL model. The model provides feedback based on the generated path and performance metrics to optimize the policy network. S55: Repeat steps S51-S54 to generate different initial solutions or optimized paths; S56: Training ends, outputting a set of initial flow paths to represent individuals in the solution.

9. The efficient routing method for ultra-high bandwidth service flows in time-sensitive networks according to claim 8, characterized in that: In step S6, the CNC optimizes the candidate paths selected in the routing calculation using the non-dominated sorting genetic algorithm NSGA-II to select the preferred path from the sending end to the receiving end that meets the service flow requirements. Specifically, it includes the following steps; S61: First set the initialization parameters, including population size Size and NSGA-II maximum number of iterations ζ. NSGA-II Crossover rate Cr NSGA-II Variation rate Mr NGGA-II ; S62: Transform the initial solution obtained in step S5 into an initial population of NSGA-II to create a parent population P0 of size Size; S63: Sort the parent population P0 according to the non-dominance level and assign each chromosome a value equal to its non-dominance level; and set the initial number of cycles to ι = 0; S64: Create a progeny population Γ0 on P0 using selection, crossover, and mutation operators based on non-dominated levels; S65: Determine whether the constraints are met. If the constraints are met, stop and return to P1; if the constraints are not met, proceed to step S66. S66: Merge the initial parent population P0 and the offspring population Γ0 to generate a merged population R. ι ; S67: From R ι Identify the non-dominated frontiers F1, F2, ..., F κ and set i = 1; S68: When |Pι + 1| < Size, calculate the crowding distance of the solutions in F; if |P i | ∪ |F ι+1 | ≤ Size, then let P i = P ι+1 ∪ F ι+1 ; otherwise, sort F according to the crowding distance and add the least crowded Size - |P i | solutions from F i to P i . ι+1 ι+1 ι+1 | solutions; Then let i = i + 1; S69: Determine if the termination condition is met. If not, set ι = ι + 1, return to step S64, and proceed to P. ι+1 The above applies selection, crossover, and mutation operators based on non-dominated levels to create a progeny population Γ. ι+1 If satisfied, the output is the solution obtained after optimization by NSGA-II; that is, the optimal path from the sender to the receiver that meets the requirements of various service flows is selected.

Images