A method, apparatus, and electronic device for network routing and time slot allocation across time scales.

By employing cross-timescale network routing and time slot allocation methods, the problems of high computational complexity and difficulty in guaranteeing deterministic latency in 5G and TSN convergence scenarios are solved, enabling efficient deterministic transmission of dynamic service flows.

CN123124093APending Publication Date: 2026-11-17BEIJING UNIV OF POSTS & TELECOMM
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611324981.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-28
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In industrial internet scenarios where 5G and time-sensitive networks are integrated, existing technologies struggle to guarantee end-to-end deterministic latency in large-scale dynamic concurrent scenarios. Furthermore, the computational complexity of global joint scheduling explodes exponentially, making it impossible to effectively handle random arrival jitter introduced by wireless jitter.

Method used

A cross-timescale network routing and time slot allocation method is adopted. By separating macro-routing and micro-time slot allocation, the target routing path and time slot allocation parameters are generated using a preset strategy generation algorithm and a deep strategy network. Combined with global time reference phase and local network state, deterministic transmission of dynamic service flows is achieved.

Benefits of technology

It reduces computational complexity, meets the timeliness requirements of dynamic business flows, forcibly absorbs wireless jitter, and achieves absolute deterministic latency guarantee across domains and end-to-end.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123124093A_ABST
    Figure CN123124093A_ABST
Patent Text Reader

Abstract

This application provides a method, apparatus, and electronic device for cross-timescale network routing and time slot allocation. The method includes: acquiring global network state characteristics of a converged network according to a scheduling period corresponding to a first timescale; generating target routing paths for the scheduled service flow in the wired domain, along with corresponding delay budget constraint parameters and a global time reference phase, using a preset policy generation algorithm; acquiring the local network state of each hop node of the target routing path according to a scheduling period corresponding to a second timescale; and generating target time slot allocation parameters for the scheduled service flow at each hop node by combining the delay budget constraint parameters and the global time reference phase; and distributing the target time slot allocation parameters determined for each hop node to the gating configuration table of each hop node for locking. This invention can forcibly absorb and smooth out millisecond-level random arrival jitter introduced by fluctuations in the wireless access air interface, ensuring absolutely deterministic end-to-end cross-domain delay transmission.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of network scheduling technology, and more specifically, to a method, apparatus, and electronic device for network routing and time slot allocation across time scales. Background Technology

[0002] In industrial internet scenarios where 5G and Time-Sensitive Networking (TSN) converge, TSN provides microsecond-level deterministic transmission on the wired side through clock synchronization and gated queuing mechanisms. However, the inherent channel variations and queuing fluctuations of the 5G wireless air interface introduce millisecond-level random arrival jitter. This macroscopic random jitter on the wireless side and the microscopic deterministic time slots on the wired side have a natural time-domain scale contradiction, making the end-to-end determinism of cross-domain services highly susceptible to damage.

[0003] When dealing with this converged scenario, existing technologies typically attempt to jointly solve the allocation of macro-routes and micro-discrete time slots on a global scale. However, as the network size, time slot granularity, and number of concurrent service flows increase, the action space of global joint scheduling explodes exponentially, making it difficult to meet the millisecond-level online scheduling timeliness requirements of dynamic service flows.

[0004] Meanwhile, existing scheduling schemes based on shortest path or greedy strategies often ignore the fragmentation and destruction of underlying continuous idle time slot resources, causing subsequent large-scale business flows to be unable to access due to the long tail of discontinuous resources, making it difficult to deliver strict end-to-end deterministic latency guarantees in large-scale dynamic concurrency scenarios. Summary of the Invention

[0005] In view of this, the purpose of the present invention is to provide a network routing and time slot allocation method, apparatus and electronic device across time scales, so as to break through the bottleneck of exponential computational complexity faced by traditional centralized joint scheduling, and to forcibly absorb wireless jitter in an endogenous cooperative manner, thereby ensuring absolute deterministic latency across domains end to end.

[0006] Firstly, a cross-timescale network routing and time slot allocation method is provided, applicable to converged networks including wireless access domains and wired domains, comprising: According to the scheduling cycle corresponding to the first time scale, the global network state characteristics of the converged network are obtained, and based on the global network state characteristics, the target routing path of the service flow to be scheduled in the wired domain, the delay budget constraint parameters corresponding to the target routing path, and the global time reference phase are generated using a preset strategy generation algorithm. According to the scheduling period corresponding to the second time scale, the local network status of each hop node of the target routing path is obtained, and the target time slot allocation parameters of the service flow to be scheduled at each hop node are generated by combining the delay budget constraint parameters and the global time reference phase; the first time scale is in milliseconds; the second time scale is in microseconds. The target time slot allocation parameters determined by each hop node are sent to the gating configuration table of each hop node for locking.

[0007] Optionally, based on global network state characteristics, a preset strategy generation algorithm is used to generate the target routing path of the service flow in the wired domain, as well as the corresponding delay budget constraint parameters and global time reference phase, including: The global network state features are input into a pre-trained deep policy network to predict the probability distribution of the initial routing path. Forward pruning is performed on the probability distribution of the initial routing paths to filter out a set of candidate paths that meet the physical constraints; Based on the link hop count, remaining link capacity, and end-to-end latency requirements of each path in the candidate path set, the latency budget constraint parameters and global time reference phase of each candidate path are calculated. The candidate strategies in the candidate strategy combination are evaluated using a pre-constructed composite reward function to determine the target routing path and the corresponding delay budget constraint parameters and global time reference phase. Among them, the delay budget constraint parameter is used to limit the search space boundary of the time slot offset that can be allocated to each hop node for the service flow; the global time reference phase is used to align the local clock reference of each hop node.

[0008] Optionally, performing forward pruning on the probability distribution of the initial routing path includes: For each initial routing path, obtain the end-to-end latency requirements and bandwidth capacity requirements of the service flow; Retrieve the available bandwidth capacity and cumulative topology delay of each link on the initial routing path; The probability distribution value of the corresponding initial route path will be set to an invalid value and discarded if any of the following conditions are met: The remaining available bandwidth capacity of any link on the path is less than the bandwidth capacity requirement of the service flow. The cumulative topology latency of the path exceeds the end-to-end latency requirement of the service flow.

[0009] Optionally, the local network status of each hop node of the target routing path is obtained, and combined with the delay budget constraint parameters and the global time reference phase, the target time slot allocation parameters for the service flow to be scheduled at each hop node are generated, including: For each hop node on the target routing path, at each scheduling time of the second time scale, obtain the context information of the current service flow from the upstream node or the source end. The context information includes at least the target time slot offset of the service flow determined by the previous hop node, and the remaining effective delay budget updated by the upstream node according to the link processing capacity. The local time slot search window is determined by taking the global time reference phase as the origin of the local clock reference and combining it with the remaining effective time delay budget. Within the local time slot search window, the target time slot allocation parameters are obtained by constructing a local conflict graph for the corresponding node and solving for the weighted maximum independent set of the local conflict graph.

[0010] Optionally, within the local time slot search window, by constructing a local conflict graph for the corresponding node and solving for the weighted maximum independent set in the local conflict graph, the target time slot allocation parameters are obtained, including: Within the local time slot search window, each candidate time slot searched is mapped to a vertex in the local conflict graph; Based on the pre-defined multi-dimensional conflict determination logic, undirected conflict edges are constructed between vertices that have mutual exclusion relationships; The composite semantic weight value corresponding to each vertex is dynamically calculated. The composite semantic weight value is used to characterize the contribution of the selected time slot to the overall system gain and the compatibility impact on subsequent services. Based on the composite semantic weight value, a distributed forward pruning algorithm is executed on the local conflict graph to solve the weighted maximum independent set. In the weighted maximum independent set, the independent vertex with the largest composite semantic weight value is selected as the target time slot, and the target time slot allocation parameters are determined based on the target time slot.

[0011] Optionally, based on composite semantic weight values, a distributed forward pruning algorithm is performed on the local conflict graph to solve for the weighted maximum independent set, including: Sort all vertices in descending order of their weight values; The forward pruning process is performed using a greedy heuristic strategy: Select the unprocessed vertex with the highest weight in the current list as the candidate vertex; Check whether the candidate vertex has an edge conflict with the vertices already selected in the set; If there is no conflict, the candidate vertex is added to the temporary optimal time slot set, and the candidate vertex and all its adjacent vertices are removed from the local conflict graph. If a conflict exists, the candidate vertex is discarded and removed from the local conflict graph. Repeat the above pruning process until all vertices in the local conflict graph have been processed, resulting in a locally conflict-free set of independent vertices, which serves as the weighted maximum independent set.

[0012] Optionally, the composite semantic weight is calculated by jointly weighting and summing the following four components: Service priority benefit component: The external priority determined based on the service quality identifier of the service flow to be scheduled is jointly determined by the internal priority determined based on the service deadline, transmission period and frame length tuning; Queue delay tolerance component: determined based on the ratio of the difference between the preset maximum end-to-end delay budget of the entire network and the cumulative queuing delay prediction value of the service flow at the current hop node, and used to give weight compensation to urgent service flows; Micro-phase consistency score component: determined based on the reciprocal of the absolute value of the time deviation between the current candidate time slot and the global time reference phase, used to characterize the degree of alignment between time slot selection and the overall network scheduling cycle; The resource fragmentation penalty component is calculated based on the distribution changes of continuous idle intervals of local ports before and after the allocation of candidate time slots.

[0013] Optionally, the resource fragmentation penalty component, calculated based on the distribution change of continuous idle intervals of local ports before and after the allocation of candidate time slots, includes: Based on the ratio of the length of the continuous idle interval in the link port time slot bitmap to the total idle capacity, calculate the first resource fragment information entropy before allocation and the second resource fragment information entropy after allocation. The difference between the information entropy of the second resource fragment and the information entropy of the first resource fragment is used as the resource fragment penalty component.

[0014] Secondly, a cross-timescale network routing and time slot allocation device is provided, applied to a converged network including a wireless access domain and a wired domain, comprising: The first generation unit is used to obtain the global network state characteristics of the converged network according to the scheduling period corresponding to the first time scale, and based on the global network state characteristics, use a preset strategy generation algorithm to generate the target routing path of the service flow to be scheduled in the wired domain and the delay budget constraint parameters and global time reference phase corresponding to the target routing path. The second generation unit is used to obtain the local network status of each hop node of the target routing path according to the scheduling period corresponding to the second time scale, and generate the target time slot allocation parameters of the service flow to be scheduled at each hop node by combining the delay budget constraint parameters and the global time reference phase; the first time scale is in milliseconds; the second time scale is in microseconds. The distribution unit is used to distribute the target time slot allocation parameters determined by each hop node to the gating configuration table of each hop node for locking.

[0015] Thirdly, an electronic device is provided, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; A processor, when executing a program stored in memory, implements any of the methods of the first aspect.

[0016] This invention provides a cross-timescale network routing and time slot allocation method, apparatus, and electronic device. It acquires the global network state characteristics of the converged network according to a scheduling cycle corresponding to a first timescale, and based on these global network state characteristics, uses a preset strategy generation algorithm to generate the target routing path of the service flow to be scheduled in the wired domain, along with the corresponding delay budget constraint parameters and global time reference phase. According to a scheduling cycle corresponding to a second timescale, it acquires the local network state of each hop node of the target routing path, and combines the delay budget constraint parameters and global time reference phase to generate the target time slot allocation parameters for the service flow to be scheduled at each hop node. The target time slot allocation parameters determined for each hop node are then distributed to the gating configuration table of each hop node for locking. This invention, through its architecture that effectively decouples macro-routing and micro-time slot alignment, avoids the exponential explosion problem of action space encountered when jointly solving macro-routing and micro-discrete time slot allocation globally, significantly reducing computational complexity and meeting the timeliness requirements of dynamic online service flow scheduling. Meanwhile, by utilizing the precise stretching of microsecond-level wired gating time slots, it is possible to forcibly absorb and smooth out millisecond-level random arrival jitter introduced by fluctuations in the wireless access air interface, thereby delivering absolute deterministic latency guarantee across domains end to end.

[0017] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A flowchart of the cross-timescale network routing and time slot allocation method provided in an embodiment of the present invention is shown; Figure 2 This invention provides an overall architecture diagram of the 5G-TSN converged network cross-timescale collaborative scheduling framework. Figure 3 This is a schematic diagram illustrating the interaction between the deep policy network and the dynamic action mask in an embodiment of the present invention; Figure 4This is a schematic diagram illustrating the construction of a local conflict graph and the solution of the weighted maximum independent set in an embodiment of the present invention; Figure 5 This is a comparison chart of the impact of different service flow scales on network scheduling success rate in an embodiment of the present invention; Figure 6 This is a comparison chart of the impact of different service flow scales on the weighted success rate in an embodiment of the present invention; Figure 7 This diagram illustrates the structure of a cross-timescale network routing and time slot allocation device provided in an embodiment of the present invention. Figure 8 A schematic diagram of the structure of an electronic device provided in an embodiment of the present invention is shown. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0021] Firstly, a cross-timescale network routing and time slot allocation method is provided, applicable to converged networks including wireless access domains and wired domains, such as... Figure 2 As shown, this converged network is, for example, an industrial internet scenario that integrates 5G technology and TSN, where the wireless access domain is responsible for the flexible access of mobile terminals, and the wired domain is responsible for the deterministic transmission of industrial control data.

[0022] exist Figure 2 The paper presents a hierarchical, cross-scale deterministic network scheduling and routing management system. At the bottom layer is the global routing guidance layer, which acts as the "brain" of the system and is responsible for generating long-term, global strategies.

[0023] This layer connects TSN switches, computing units, and heterogeneous access units such as 5G networks including DS-TT and NW-TT, and covers various terminal devices such as robots and sensors. Its core function lies in state awareness and policy generation: by collecting information such as the remaining capacity, latency, and fragmentation entropy of the entire network links, it runs algorithms such as submodal optimization or gated neural networks to calculate the globally optimal routing path, latency budget constraints, and global time reference phase.

[0024] The global layer does not directly process data packets, but instead generates "policy boundaries" or "hard pruning constraints" containing these key parameters, providing explicit navigation instructions to the lower layers.

[0025] At the top is the distributed execution layer, which acts as the system's "hands and feet," responsible for distributing and executing specific tasks from a short-cycle, local perspective.

[0026] This layer directly interacts with the servers, storage units, and data flow processing. The table at the top of the diagram shows the specific scheduling plan (Sche), which includes the business flow identifier. and offset parameters on different links or time slots (such as) , ), indicating the error tolerance or a specific time slot marker ( , ).

[0027] The execution layer receives macro-level policies from the global layer and translates them into microsecond-level specific scheduling instructions, such as precisely determining which port a data packet will be sent from at a specific microsecond. By executing these micro-level instructions independently or collaboratively on each node, the global deterministic requirement is achieved by ensuring that data flows are transmitted strictly according to the predetermined path and timing.

[0028] like Figure 1 As shown, the method includes the following steps: Step S101: According to the scheduling period corresponding to the first time scale, obtain the global network state characteristics of the converged network, and based on the global network state characteristics, use the preset strategy generation algorithm to generate the target routing path of the service flow to be scheduled in the wired domain, as well as the delay budget constraint parameters and global time reference phase corresponding to the target routing path.

[0029] The primary timescale is in the millisecond range. In 5G-TSN converged scenarios, the radio air interface is affected by channel fading, interference, and queuing fluctuations, resulting in arrival jitter typically in the millisecond range. Therefore, a millisecond-level scheduling cycle is used to address the macroscopic statistical characteristics of dynamic service flow arrivals and to perform global-level resource rectification and path planning.

[0030] Global network state characteristics include node connectivity, link capacity, time slot occupancy status, and multi-dimensional physical characteristics of heterogeneous service flows.

[0031] The generated target routing path specifies the sequence of wired domain nodes that the service flow passes through from the source to the destination. The delay budget constraint parameter is used to limit the search space boundary of the time slot offset that the service flow can be allocated at each hop node, that is, to specify the maximum queuing time that the service flow is allowed to consume in the wired domain, so as to reserve enough buffer to absorb random jitter on the radio side.

[0032] The global time reference phase serves as a flexible time anchor point, used to align the local clock reference of each hop node, guiding periodic services to implicitly align to a unified macroscopic time domain reference system, thereby mitigating the accumulated jitter time amplified across hops.

[0033] Step S102: According to the scheduling period corresponding to the second time scale, obtain the local network status of each hop node of the target routing path, and combine the delay budget constraint parameters and the global time reference phase to generate the target time slot allocation parameters of the service flow to be scheduled in each hop node.

[0034] The second time scale is at the microsecond level. The wired side of the TSN implements a microsecond-level gated queuing mechanism through a Time-Aware Shaper (TAS), which requires extremely high time precision. Therefore, the microsecond-level scheduling cycle is used to handle the local queuing behavior of a single service flow at a specific edge node.

[0035] In this step, each hop node no longer performs a global search, but instead searches for the optimal transmission slot only within a local slot search window defined by the delay budget constraint parameter, combined with the starting reference determined by the global time reference phase.

[0036] Step S103: Send the target time slot allocation parameters determined by each hop node to the gating configuration table of each hop node for locking.

[0037] Specifically, the target timeslot allocation parameters include the specific transmission timeslot offset of the service flow at each hop switch or gateway node. Once determined, these parameters are written into the gate control list (GCL) of the underlying switching node.

[0038] The gating configuration table controls the opening and closing time of the queue gates, ensuring that service frames are sent at precise microsecond intervals, thereby eliminating queuing conflicts and guaranteeing deterministic transmission.

[0039] Before issuing the lock, an end-to-end deterministic check can be performed. The total end-to-end delay is calculated based on the target time slot offset of each hop lock. The formal issuance is allowed only if the total delay is not greater than the service deadline; otherwise, rescheduling is triggered.

[0040] This embodiment, through its architecture that effectively decouples macroscopic routing from microscopic time slot alignment, avoids the exponential explosion of action space encountered when jointly solving macroscopic routing and microscopic discrete time slot allocation on a global scale. This significantly reduces computational complexity and meets the timeliness requirements of dynamic service flow online scheduling. Simultaneously, by utilizing the fine stretching of microsecond-level wired gated time slots, it can forcibly absorb and smooth out millisecond-level random arrival jitter introduced by fluctuations in the wireless access air interface, thereby ensuring absolute deterministic latency across domains from end to end.

[0041] Based on the above embodiments, and based on global network state characteristics, a preset policy generation algorithm is used to generate the target routing path of the service flow in the wired domain, as well as the delay budget constraint parameters and global time reference phase corresponding to the target routing path, including: Step S101A: Input the global network state features into the pre-trained deep policy network to predict the probability distribution of the initial routing path.

[0042] like Figure 3 As shown, the global routing guidance module, as the macro decision-making center, first collects the network topology connection relationship, instantaneous queuing load of nodes, and allocable bandwidth of heterogeneous links in real time through the southward telemetry interface of the control plane.

[0043] These multidimensional data are matrix-encoded into initial node feature matrices and edge feature matrices of a graph structure.

[0044] By using a multi-layer graph convolutional neural network (GCN) to aggregate multi-hop neighbor features, local random noise caused by dynamic changes in the wireless channel is filtered out, and a low-dimensional global environmental state representation is extracted and output.

[0045] The global environment state is input into a pre-trained deep policy network, where the value network predicts a global routing policy probability distribution based on the current hidden state, representing the probability that each physical candidate route is selected as the optimal deterministic path.

[0046] Specifically, in this embodiment, the solution process of the deep policy network is modeled as a joint optimization problem P1:

[0047] In the formula, and the target path matrix to be decided; The first indicator variable characterizing the optimal conflict-free time slot offset vector of the service flow at each hop node; This represents the set of service flows currently awaiting scheduling. This is one of the business flows; This is a variable indicating successful scheduling. and Rewards are given based on priority and latency margin. For link The resource fragmentation penalty term is calculated using the idle time slot information entropy:

[0048]

[0049] in, For link The first port time slot bit diagram The length of a card slot in a continuous free interval. This represents the total free capacity. Higher information entropy indicates that contiguous resource blocks are fragmented more finely, leading to more severe scheduling degradation.

[0050] The joint optimization problem P1 must be subject to four rigid physical constraints: Condition 1, Global Jitter Constraint: The random time-domain transmission variance of the wireless segment must be forcibly smoothed to satisfy... .

[0051] Deterministic transmission jitter within the TSN domain; : Latency fluctuation range of 5G wireless link (maximum value minus minimum value); : Jitter threshold.

[0052] Condition 2, Time Slot Uniqueness Constraint: Prevent collisions between card slots in the same domain on the same physical link, satisfying... .

[0053] On the same physical link Above, any two different flows and The time slots cannot overlap.

[0054] :flow In the link The initial time slot offset.

[0055] :flow The cycle (if it is a periodic business).

[0056] , Integer index, indicating the first... One cycle.

[0057] :flow Required time slot length (bandwidth requirement).

[0058] It should be noted that this formula assumes the business is cyclical.

[0059] Condition 3, Sequential Consistency Constraint: Maintain the hop-by-hop causal sequence and processing overhead of the cascaded path, satisfying... .

[0060] It should be ensured that data packets require a propagation delay from the previous hop to the next hop. and processing latency Plus safety margin : : Current jump; : Next jump time.

[0061] Condition 4, Deadline Constraint: Fulfill the end-to-end hard QoS latency red line and meet the requirements. .

[0062] Meaning: The transmission completion time of the last node cannot exceed the deadline of the service flow. .

[0063] : Business arrival time.

[0064] The starting offset of the scheduling on the last link.

[0065] : Transmission / processing time of the last link segment.

[0066] Because the decision variables in the joint optimization problem P1 are highly nonlinearly coupled, the solution space explodes exponentially in dimensionality as the flow expands (belonging to a strong NP-hard problem).

[0067] This invention decouples the cross-scale problem into two sub-problems: macro-level strategy generation on a long timescale and micro-level time-slot resolution on a short timescale. This not only greatly reduces the search dimension of the model solution, but also achieves precise and efficient optimization of dynamic service provisioning configuration through inter-layer prediction, decision-making, and closed-loop entropy pruning feedback.

[0068] Furthermore, to transform the continuous probability distribution into a discrete, physically feasible path selection decision, this embodiment introduces a temperature scaling mechanism. Given an unscaled routing strategy probability distribution, a temperature parameter is introduced. (generally The probability distribution is modified using the following formula:

[0069] in, For the first The original probability values ​​of the candidate paths; For the first The original probability values ​​of the candidate paths; These are the scaled probability values. Temperature scaling helps smooth the probability distribution and avoids premature convergence to local optima.

[0070] Step S101B: Perform forward pruning on the probability distribution of the initial routing paths to filter out a set of candidate paths that meet the physical constraints.

[0071] In one feasible implementation, performing forward pruning on the probability distribution of the initial routing path includes: Step A: For each initial routing path, obtain the end-to-end latency requirements and bandwidth capacity requirements of the service flow.

[0072] Step B: Retrieve the available bandwidth capacity and cumulative topology delay of each link on the initial routing path.

[0073] Step C: When any of the following conditions are met, set the probability distribution value of the corresponding initial routing path to an invalid value and discard it: The remaining available bandwidth capacity of any link on the path is less than the bandwidth capacity requirement of the service flow. The cumulative topology latency of the path exceeds the end-to-end latency requirement of the service flow.

[0074] Specifically, this embodiment uses the Entropy-based Dynamic Action Mask (EDAM) algorithm to implement the aforementioned forward hard masking pruning.

[0075] Based on the frame length of the service flow to be accessed Link nominal capacity and discrete time slot length Calculate the number of consecutive time slots required for each hop of the service flow on the link. .

[0076] Simultaneously, under the cross-queue mutual exclusion view, the longest available continuous idle time slot block length exists for each link on the current candidate path. If the required number of consecutive time slots is greater than the longest consecutive free block length of any link in the candidate path (i.e., ... If the lower bound of the pure physical topology delay of the path exceeds the service flow deadline, or if it contains a temporarily unreachable link due to a temporary failure, the action space masking mechanism directly triggers forward hard masking, forcibly correcting the policy exploration state evaluation value or selection probability corresponding to the infeasible path to a minimum value (e.g., negative infinity). ).

[0077] In this way, the exponentially expanding path search space is dynamically compressed into a simplified subset that is absolutely physically feasible, significantly improving search efficiency.

[0078] Step S101C: Based on the link hop count, remaining link capacity, and end-to-end latency requirements of each path in the candidate path set, calculate the latency budget constraint parameters and global time reference phase of each candidate path to obtain the candidate strategy combination.

[0079] To efficiently explore the routing space in complex network topologies, discrete link selections need to be transformed into structured policy representations.

[0080] Specifically, through one-hot coding technology, each candidate path is transformed into a second indicator variable representing whether a node and link are selected into the target physical path. (i.e., candidate binary routing matrix), thereby establishing the basic set of routing candidates.

[0081] Based on the aforementioned second indicator variable By invoking a low-complexity graph-level linear constraint allocator, the third indicator variable corresponding to the budget tightness factor for the time delay is further derived. and the fourth indicator variable indicating the phase of the global time reference. This forms multiple sets of relaxed control plane candidate policy combinations. .

[0082] Third indicator variable Used to dynamically limit the upper limit boundary of the underlying local search window, ensuring that business latency meets the constraints.

[0083] The fourth indicator variable is the global time reference phase. As a flexible time anchor, it guides periodic business to implicitly align with a unified macroscopic time domain reference system to resolve cross-link time synchronization deviations.

[0084] Step S101D: Use the pre-built composite reward function to evaluate each candidate policy in the candidate policy combination, and determine the target routing path and the corresponding delay budget constraint parameters and global time reference phase.

[0085] The composite reward function is shaped into a nonlinear weighted combination of network physical indicators, including at least a basic scheduling success reward, a service priority guarantee reward based on service quality identifier and expected deadline, a latency guarantee reward to maximize end-to-end remaining time margin, and a resource fragmentation penalty.

[0086] The resource fragmentation penalty term is positively correlated with the increase in idle time slot fragment entropy in the underlying feedback, forcing the agent to actively avoid paths that are prone to causing time slot discontinuity and long-tail breakage during macro-path planning. Finally, the path with the highest score in the composite reward function is selected as the target routing path.

[0087] The latency budget constraint parameter reflects the maximum queuing time allowed for a service flow in the wired forwarding segment of the target path.

[0088] After obtaining the candidate strategy combinations, the next core task is to select the globally optimal solution from the many candidates and drive the continuous evolution of the system.

[0089] This embodiment introduces an evaluator based on submodular optimization scoring to accurately evaluate candidate strategy combinations. This evaluation process aims to maximize the objective function of the global bootstrapping configuration problem P2, which is mathematically expressed as:

[0090]

[0091]

[0092] The core driving force of the evaluator comes from its built-in single-step compound reward function: ; In the formula To preset the balance coefficient, Rewards for successful basic scheduling. To ensure the reward items are prioritized for business operations, As a reward item to ensure time delay margin, This is a penalty for resource fragmentation.

[0093] Based on the first indicator variable corresponding to the current time slot Based on the backfeedback conflict indicators and predicted resource fragmentation, the evaluator uses a submodular optimization algorithm to select the final optimal solution: the objective's second indicator variable. The third indicator variable of the objective and the fourth indicator variable of the target .

[0094] This not only determined the final target routing path, but also locked the precise delay budget constraints and global time reference phase required for that path.

[0095] After determining the optimal strategy, the execution and learning phase begins.

[0096] At the start of the next microsecond-level scheduling cycle, the updated global policy boundary is used as a hard pruning constraint for actions to quickly solve the global problem P1, thereby updating the first indicator variable of the micro-slot offset. This completes the closed loop of this scheduling process.

[0097] Meanwhile, in order to achieve adaptive improvement in long-term performance, the environmental characteristics of this microsecond-level time slot (including the first indicator variable) will be incorporated. ) and the determined second indicator variable to construct sample pairs The sample is used to train a gated neural network and initiate a new cross-scale asynchronous action iteration program to continuously optimize the routing decision-making capability of subsequent time slots.

[0098] The global time reference phase serves as a flexible time anchor point, used to align the local clock reference of each hop node.

[0099] Because the clocks of different switches in a TSN network may have slight deviations, and periodic services require a unified time base, the global time reference phase is calculated and distributed uniformly by the central controller. This guides periodic services to implicitly align with a unified macroscopic time domain reference system, mitigating the accumulated jitter time amplified by cascading connections. In this way, the consistency of network-wide scheduling in the time dimension is ensured.

[0100] Based on the above embodiments, the local network status of each hop node of the target routing path is obtained, and combined with the delay budget constraint parameters and the global time reference phase, the target time slot allocation parameters for the service flow to be scheduled at each hop node are generated, including: Step S102A: For each hop node on the target routing path, at each scheduling time of the second time scale, obtain the context information of the current service flow from the upstream node or the source.

[0101] The context information includes at least the target timeslot offset of the service flow determined by the previous hop node, and the remaining effective delay budget updated by the upstream node based on the link processing capacity.

[0102] like Figure 4 As shown, on a short timescale of microseconds, the distributed execution layer (such as a TSN switch) is responsible for performing local time slot alignment. First, the node receives the global bootstrapping strategy from the control plane, including the global time reference phase and delay budget constraint parameters.

[0103] Meanwhile, nodes obtain the context information of the current service flow from upstream nodes or the source. The "target time slot offset determined by the previous hop node" in the context information is used to ensure the consistency of hop-by-hop causal timing, that is, the current node's transmission start time must be later than the previous hop's transmission end time plus the link propagation delay and processing delay.

[0104] The remaining effective delay budget is a value that is dynamically updated by the upstream node based on the actual queuing situation. It reflects the maximum queuing delay that the business flow can tolerate in subsequent links.

[0105] Step S102B: Using the global time reference phase as the origin of the local clock reference, determine the local time slot search window by combining the remaining effective delay budget.

[0106] Based on the above information, the node determines the local time slot search window, using the global time reference phase. Using the local clock as the reference origin and combining the remaining effective delay budget, a time-bounded search interval is defined.

[0107] For example, if the timestamp of the service frame arriving at the current node is The remaining effective delay budget is (in, As the time-delay budget tightness factor, Given the end-to-end deadline requirement of the current business flow, the upper bound of the local slot search window is... It can be represented as Within this window, the system operates in fixed microsecond steps. Iterate through all possible candidate time slot offsets.

[0108] Step S102C: Within the local time slot search window, the target time slot allocation parameters are obtained by constructing the local conflict graph of the corresponding node and solving the weighted maximum independent set of the local conflict graph.

[0109] In one feasible implementation, within a local time slot search window, by constructing a local conflict graph for the corresponding nodes and solving for the weighted maximum independent set of the local conflict graph, the target time slot allocation parameters are obtained, including: Step S102C1: Within the local time slot search window, map each candidate time slot being searched to a vertex in the local conflict graph.

[0110] Specifically, local conflict diagram It is a model that formalizes the resolution of temporal collisions into graph theory constraints.

[0111] Among them, the vertex set Each vertex in Represents the offset of a candidate time slot within the search window. Edge set An undirected edge in the diagram indicates that there is a mutual exclusion relationship between two candidate time slots, and they cannot be selected at the same time.

[0112] Step S102C2: Based on the preset multi-dimensional conflict determination logic, construct undirected conflict edges between vertices that have mutual exclusion relationships.

[0113] The multi-dimensional conflict determination logic is used to determine any two candidate time slots. and Whether there is a conflict between them, the criteria for determination include but are not limited to: 1. Time overlap: Two time slots partially or completely overlap on the time axis; 2. Cross-queue mutual exclusion: Within the same period, two time slots violate the exclusivity rules between queues of different priorities; 3. Insufficient protection interval: The time interval between two time slots is less than the minimum protection interval required by the physical layer; 4. Causal timing disruption: Choosing a particular time slot can cause the total end-to-end delay to exceed the deadline, or disrupt hop-by-hop causal consistency on the cascaded path.

[0114] If any of the above conditions are met, then at the vertex and Construct an undirected conflict edge between them.

[0115] Step S102C3: Dynamically calculate the composite semantic weight value corresponding to each vertex.

[0116] The composite semantic weight value is used to characterize the contribution of the selected vertex to the overall system gain and the compatibility impact on subsequent services.

[0117] In this embodiment, the composite semantic weight is calculated by jointly weighting and summing the following four components: Service priority benefit component: The external priority determined based on the service quality identifier of the service flow to be scheduled is jointly determined by the internal priority determined based on the service deadline, transmission period and frame length tuning; Queue delay tolerance component: determined based on the ratio of the difference between the preset maximum end-to-end delay budget of the entire network and the cumulative queuing delay prediction value of the service flow at the current hop node, and used to give weight compensation to urgent service flows; Micro-phase consistency score component: determined based on the reciprocal of the absolute value of the time deviation between the current candidate time slot and the global time reference phase, used to characterize the degree of alignment between time slot selection and the overall network scheduling cycle; The resource fragmentation penalty component is calculated based on the distribution changes of continuous idle intervals of local ports before and after the allocation of candidate time slots.

[0118] Specifically, composite semantic weight value The calculation comprehensively considers business priority, latency urgency, phase consistency, and resource fragmentation penalties. The specific calculation formula is as follows:

[0119] in, These are the preset weighted balance coefficients.

[0120] The service priority benefit component is jointly determined by the external priority based on the service quality identifier of the service flow to be scheduled and the internal priority based on the service deadline, transmission period and frame length tuning. Among them, the business priority revenue component It integrates static configuration and dynamic requirements. External priority is directly mapped to the 5G QoS identifier (5QI) or TSN flow category to ensure that high-value services receive basic weight advantages; internal priority is dynamically adjusted according to the urgency of service flows. For example, for short-cycle, small-frame services that are close to their deadline, their internal priority weight is increased to prevent them from being squeezed out by long-cycle, large-frame services due to their short frame length.

[0121] The queue delay tolerance component is determined based on the ratio of the difference between the preset maximum end-to-end delay budget for the entire network and the predicted cumulative queuing delay of the service flow at this hop node. It is used to provide weight compensation for urgent service flows and satisfies the mapping formula: In the formula The deadline for the business flow. This refers to the currently accumulated fixed delay items (including propagation delay, processing delay, and queuing delay that have already occurred). Indicates an indicator variable, when the business flow The value is 1 when the device is active, and 0 otherwise. when near When the denominator remains unchanged, the numerator decreases. However, this component is usually designed as an inverse ratio or a difference ratio to reflect urgency. If defined as a difference ratio, the less time remaining, the smaller this component may be. Therefore, in actual weighting, the reciprocal or normalized remaining ratio is often used as a positive incentive, or, as in this embodiment, a negative penalty mechanism or adjustment coefficient is used. This ensures that urgent business processes receive higher weight. More preferably, this weight can be designed to be positively correlated with the remaining latency budget, meaning that the less the remaining budget, the higher the weight compensation, in order to prioritize business processes that are about to time out.

[0122] The micro-phase consistency score component is determined based on the reciprocal of the absolute value of the time deviation between the current candidate time slot and the global time reference phase. It is used to characterize the alignment between time slot selection and the overall network scheduling cycle. For example, the micro-phase consistency score component can utilize the distributed global reference phase as a distributed flexible anchor point, and measure the alignment between the candidate time slot and the global phase using a cosine similarity nonlinear calculation formula.

[0123] in, For candidate time slot offsets, As a global time reference phase, This is for traffic over-cycle. When the candidate time slot is perfectly aligned with the global time reference phase, the cosine value is 1, resulting in the highest score; the greater the deviation, the lower the score. This helps to form a neat traffic waveform on a macroscopic level, reducing the overall network jitter accumulation.

[0124] The resource fragmentation penalty component is calculated based on the distribution changes of continuous idle intervals of local ports before and after the allocation of candidate time slots.

[0125] The resource fragmentation penalty component is a key feature that distinguishes this invention from traditional greedy algorithms. Traditional algorithms often blindly select the earliest available time slot, resulting in consecutive idle blocks being fragmented into unusable small pieces. This embodiment quantifies the degree of resource fragmentation and applies a penalty to the time slot selection that exacerbates fragmentation.

[0126] Furthermore, the resource fragmentation penalty component, calculated based on the distribution changes of continuous idle intervals of local ports before and after the allocation of candidate time slots, includes: Step 1: Based on the ratio of the length of the continuous idle interval in the link port time slot bitmap to the total idle capacity, calculate the first resource fragment information entropy before allocation and the second resource fragment information entropy after allocation.

[0127] Specifically, the system maintains a bitmap reflecting the local port timeslot occupancy status. This is used when evaluating a candidate timeslot. First, a tentative virtual occupancy is performed, assuming the time slot has already been allocated. Then, the length distribution of all consecutive free intervals (Gap) on the port before and after occupancy is statistically analyzed. Assume there are a total of [number missing] free intervals on the port. The nth consecutive free interval, the nth The length of each consecutive free interval is The total free capacity of the port is Then the information entropy of resource fragments. Defined as:

[0128] This formula draws on the Shannon entropy concept from information theory. When idle time is evenly distributed across multiple small intervals, the entropy value is high, indicating highly fragmented resources; when idle time is concentrated in a few large intervals, the entropy value is low, indicating good resource continuity, which is beneficial for large-frame service access.

[0129] Step 2: Use the difference between the information entropy of the second resource fragment and the information entropy of the first resource fragment as the resource fragment penalty component.

[0130] Resource fragment penalty weight The entropy of the second resource fragment after exploratory occupancy is calculated. Information entropy of the first resource fragment before it was occupied Difference:

[0131] like This indicates that the allocation of this time slot leads to more dispersed idle intervals and increased fragmentation, which will consequently reduce the composite semantic weight of this vertex; if This indicates that the allocation of this time slot actually integrates idle resources (e.g., fills the gap between two small fragments and merges them into a large contiguous block), reducing fragmentation. The system will then provide positive incentives or reduce penalties.

[0132] For example, suppose a port currently has two consecutive free blocks, each 10 microseconds long, for a total of 20 microseconds of free time. The probability is 0.5 for both, and the entropy is... .

[0133] Choosing a time slot in the middle for allocation, cutting or splitting two 10-microsecond free blocks, may result in more smaller free blocks and an increase in entropy. Conversely, choosing a time slot adjacent to an existing free block may not significantly change the distribution, or under certain boundary conditions, may reduce entropy by eliminating extremely small fragments.

[0134] By minimizing The algorithm tends to select time slots that can maintain or improve the continuity of idle resources, thereby reserving a sufficient continuous transmission window for large data frames that may arrive later, fundamentally improving the long-term schedulability and robustness of the network under high load.

[0135] Step S102C4: Based on the composite semantic weight value, perform a distributed forward pruning algorithm on the local conflict graph to solve the weighted maximum independent set. In the weighted maximum independent set, select the independent vertex with the largest composite semantic weight value as the target time slot, and determine the target time slot allocation parameters based on the target time slot.

[0136] In one feasible implementation, a distributed forward pruning algorithm is performed on the local conflict graph based on composite semantic weight values ​​to solve for the weighted maximum independent set, including: Step S1: Sort all vertices in descending order of their weight values.

[0137] Step S2: Perform the forward pruning process using a greedy heuristic strategy.

[0138] Step S3: Select the unprocessed vertex with the highest weight in the current list as a candidate vertex.

[0139] Step S4: Check whether the candidate vertex has an edge conflict with the vertices currently selected in the set.

[0140] Step S5: If there is no conflict, add the candidate vertex to the temporary optimal time slot set and remove the candidate vertex and all its adjacent vertices in the local conflict graph.

[0141] Step S6: If a conflict exists, discard the candidate vertex and remove it from the local conflict graph. Repeat the above pruning process until all vertices in the local conflict graph have been processed, resulting in a locally conflict-free set of independent vertices, which serves as the weighted maximum independent set.

[0142] During the actual execution process, the system initializes an empty temporary optimal time slot set. First, sort all vertices in descending order of their weights and store them in a list. .

[0143] Then, from Extract the first vertex from the middle .examine Whether or not Any vertex in the array has a conflicting edge. If there is no conflict, then... join in And from the figure Remove from and all its neighboring vertices (because neighboring vertices are...) (Conflict, therefore not selectable).

[0144] If there is a conflict, proceed directly from the diagram. Remove from .

[0145] Repeat this process until a list is reached. Empty or graph There are no remaining vertices. The final result is... This is the approximately optimal weighted maximal independent set.

[0146] The time slot corresponding to the vertex with the highest weight (usually the first vertex selected, or the weights are compared again in the set) is selected as the target time slot. Based on the target time slot and the global reference phase, target time slot allocation parameters, such as the transmission time slot offset, can be generated.

[0147] Traditional integer linear programming solutions to NP-hard problems require exponential time complexity, while the greedy heuristic strategy employed in this application reduces the computational complexity to zero through sorting and local pruning. or Levels (depending on the sorting algorithm and conflict checking implementation), where The number of candidate vertices is denoted by . This allows the method to complete the solution within a microsecond-level scheduling cycle, meeting the real-time requirements of online scheduling.

[0148] By transforming the microscopic time slot collision problem into a graph theory problem and solving it with a low-complexity algorithm, this embodiment not only avoids the dimensional explosion of global joint scheduling, but also guides time slot selection through composite semantic weights, taking into account service priority, latency guarantee and resource continuity, thereby improving the long-term utilization efficiency of network resources while ensuring deterministic transmission.

[0149] In another embodiment, the start time slot of the first hop and the end time slot of the last hop can be known based on the locked target time slot offsets for each hop, and the scheduling time span can be calculated by subtracting them. ; The pure topology transmission time is obtained by linearly summing the link propagation delay, switching processing delay, and frame transmission time. The total end-to-end delay is obtained by combining and summing the results. .

[0150] If and only if When the verification constraints are met, the system generates a submission confirmation and allows the scheduling parameters to be officially sent and written to the execution; if the verification fails, the current temporary parameters are discarded and a rescheduling is triggered at the upper layer.

[0151] In some embodiments, this application also provides a cross-timescale policy iterative update mechanism for achieving adaptive evolution of the global deep policy network. This mechanism calculates the resource fragmentation entropy increment and local conflict flag caused by local queuing behavior based on the telemetry feedback from the underlying network and the first indicator variable corresponding to the current time slot. It then uses a semantic-first experience replay technique incorporating service quality awareness to adjust the backsampling weight distribution of the deep policy network, iteratively optimizing the global network parameters in a closed-loop manner.

[0152] Specifically, after scheduling is completed on a short timescale of microseconds, the distributed execution layer will transmit the scheduling results and status information back to the feedback closed-loop update module of the control plane.

[0153] The feedback signals mainly include two types: first, a local conflict flag, which is forcibly triggered and set to a valid value when the underlying node cannot find a conflict-free time slot for a high-priority service flow within the limited latency budget, resulting in time slot allocation failure or severe queuing congestion; second, the resource fragmentation entropy increment, which is the difference in information entropy between the distribution of continuous idle intervals of local ports before and after the tentative occupancy calculated in the previous embodiment. These two parameters together characterize the adaptability and resource destruction degree of the current global guidance strategy at the micro-execution level.

[0154] Whenever a millisecond-level long time scale corresponding to the global rectified transmission characteristics of dynamic concurrent service flows or a specific policy iteration update cycle is reached, the system activates the service quality-aware semantic-first experience replay technology.

[0155] Traditional experience replay mechanisms typically sample historical state transitions uniformly, making it difficult for the model to learn from rare, extreme congestion scenarios. This embodiment addresses this problem by reconstructing the sampling weight distribution of the experience replay pool.

[0156] Specifically, the local conflict flag is non-linearly and exponentially coupled with the original priority attribute of the business flow as an additional semantic weighting factor. Samples that have experienced local conflicts and involve high-priority business flows are assigned extremely high sampling weights; samples whose resource fragment entropy increment exceeds a preset threshold are also given a higher probability of being sampled.

[0157] During the training of deep policy networks, this reconstructed sampling weight distribution forces the agent to extremely skew its fit to the overloaded and congested samples that lead to long-tail conflicts, severe fragmentation, and failure of high-quality flow allocation when updating gradients during backpropagation.

[0158] This means that the model no longer simply pursues the average scheduling success rate, but is forced to learn how to avoid path selections that seem to have high short-term gains but will lead to resource shortages or congestion and collapse in the long run.

[0159] In this way, the global deep policy network gains powerful congestion prediction and dynamic self-healing capabilities, enabling it to proactively avoid congestion-prone topology areas in subsequent global routing guidance and generate more robust delay budget constraint parameters.

[0160] To verify the effectiveness of this feedback loop mechanism, a comparative experiment was conducted in this application.

[0161] like Figure 5 As shown, under different service flow load conditions, with the increase of network load rate, the scheduling success rate of traditional greedy algorithm and baseline reinforcement learning algorithm shows a cliff-like drop, especially in the high load range where the load rate exceeds 80%, due to the accumulation of resource fragmentation and the inability of local congestion to heal itself, the scheduling success rate is significantly reduced.

[0162] In contrast, the scheme employing the cross-timescale collaborative scheduling and feedback closed-loop update method proposed in this application maintains a high scheduling success rate even under extremely high loads exceeding 90%. This is because the feedback mechanism enables the global policy to promptly perceive the underlying resource degradation state and quickly adjust routing preferences through experience replay, avoiding the repeated exploration of invalid paths.

[0163] In addition, such as Figure 6 As shown, under the influence of different priority service flow loads, traditional methods often fail to guarantee the scheduling priority of high-priority services during congestion due to a lack of fine-grained awareness of priority and conflict coupling. This application, however, ensures the scheduling priority of high-priority services under extreme conditions through service quality-aware semantic priority experience replay.

[0164] Data shows that in mixed load scenarios where the proportion of high-priority services increases, the scheduling success rate of this application decreases by a much smaller margin than that of the comparative scheme, demonstrating the advantages of this closed-loop architecture in ensuring deterministic transmission of critical services.

[0165] This mechanism, which is based on micro-feedback-driven macro-strategy evolution, has achieved a leap from static rule scheduling to dynamic adaptive intelligent scheduling, effectively solving the end-to-end deterministic guarantee problem of 5G-TSN converged networks in large-scale dynamic concurrent scenarios.

[0166] Based on the same inventive concept, this invention provides a cross-timescale network routing and time slot allocation device, applicable to converged networks including wireless access domains and wired domains, such as... Figure 7 As shown, it includes: The first generation unit 701 is used to obtain the global network state characteristics of the converged network according to the scheduling period corresponding to the first time scale, and based on the global network state characteristics, use a preset strategy generation algorithm to generate the target routing path of the service flow to be scheduled in the wired domain and the delay budget constraint parameters and global time reference phase corresponding to the target routing path. The second generation unit 702 is used to obtain the local network status of each hop node of the target routing path according to the scheduling period corresponding to the second time scale, and generate the target time slot allocation parameters of the service flow to be scheduled at each hop node by combining the delay budget constraint parameters and the global time reference phase; the first time scale is in the millisecond level; the second time scale is in the microsecond level. The sending unit 703 is used to send the target time slot allocation parameters determined by each hop node to the gating configuration table of each hop node for locking.

[0167] Based on the same technical concept, embodiments of the present invention also provide an electronic device, such as... Figure 8 As shown, it includes a processor 801, a communication interface 802, a memory 803, and a communication bus 804, wherein the processor 801, the communication interface 802, and the memory 803 communicate with each other through the communication bus 804.

[0168] Memory 803 is used to store computer programs; When the processor 801 executes the program stored in the memory 803, it implements the steps of the above method embodiments.

[0169] The communication bus mentioned in the above electronic devices can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of illustration, only one thick line is used to represent it in the diagram, but this does not mean that there is only one bus or one type of bus.

[0170] The communication interface is used for communication between the aforementioned electronic devices and other devices.

[0171] The memory may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0172] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0173] The network routing and time slot allocation device across time scales provided in this embodiment of the invention can be specific hardware on the device or software or firmware installed on the device. The implementation principle and technical effects of the device provided in this embodiment of the invention are the same as those in the foregoing method embodiments. For the sake of brevity, any parts not mentioned in the device embodiments can be referred to the corresponding content in the foregoing method embodiments. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can all be referred to the corresponding processes in the above method embodiments, and will not be repeated here.

[0174] In the embodiments provided by this invention, it should be understood that the disclosed apparatus and method can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the displayed or discussed mutual couplings, direct couplings, or communication connections may be through some communication interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms.

[0175] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0176] In addition, the functional units in the embodiments provided by the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0177] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0178] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. In addition, the terms "first", "second", "third", etc. are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0179] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention. All should be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A cross-timescale network routing and time slot allocation method, applied to a converged network including a wireless access domain and a wired domain, characterized in that, include: According to the scheduling period corresponding to the first time scale, the global network state characteristics of the converged network are obtained, and based on the global network state characteristics, the target routing path of the service flow to be scheduled in the wired domain and the delay budget constraint parameters and global time reference phase corresponding to the target routing path are generated using a preset strategy generation algorithm. According to the scheduling period corresponding to the second time scale, the local network status of each hop node of the target routing path is obtained, and the target time slot allocation parameters of the service flow to be scheduled in each hop node are generated by combining the delay budget constraint parameters and the global time reference phase. The first time scale is in the millisecond range; The second time scale is on the microsecond level; The target time slot allocation parameters determined by each hop node are sent to the gating configuration table of each hop node for locking.

2. The method according to claim 1, characterized in that, Based on the global network state characteristics, a preset policy generation algorithm is used to generate the target routing path of the service flow in the wired domain, as well as the delay budget constraint parameters and global time reference phase corresponding to the target routing path, including: The global network state features are input into a pre-trained deep policy network to predict the probability distribution of the initial routing path; Forward pruning is performed on the probability distribution of the initial routing path to filter out a set of candidate paths that meet the physical constraints; Based on the link hop count, remaining link capacity, and end-to-end latency requirements of the service flow for each path in the candidate path set, the latency budget constraint parameters and global time reference phase of each candidate path are calculated to obtain the candidate strategy combination. The candidate strategies in the candidate strategy combination are evaluated using a pre-constructed composite reward function to determine the target routing path and the corresponding delay budget constraint parameters and global time reference phase. The delay budget constraint parameter is used to limit the search space boundary of the time slot offset that can be allocated to each hop node for the service flow; the global time reference phase is used to align the local clock reference of each hop node.

3. The method according to claim 2, characterized in that, Performing forward pruning on the probability distribution of the initial routing path includes: For each initial routing path, obtain the end-to-end latency requirements and bandwidth capacity requirements of the service flow; Retrieve the available bandwidth capacity and cumulative topology delay of each link on the initial routing path; The probability distribution value of the corresponding initial route path will be set to an invalid value and discarded if any of the following conditions are met: The remaining available bandwidth capacity of any link on the path is less than the bandwidth capacity requirement of the service flow. The cumulative topology latency of the path exceeds the end-to-end latency requirement of the service flow.

4. The method according to claim 1, characterized in that, The local network status of each hop node of the target routing path is obtained, and combined with the delay budget constraint parameters and the global time reference phase, the target time slot allocation parameters for the service flow to be scheduled at each hop node are generated, including: For each hop node on the target routing path, at each scheduling moment of the second time scale, obtain the context information of the current service flow from the upstream node or the source end. The context information includes at least the target time slot offset of the service flow determined by the previous hop node, and the remaining effective delay budget updated by the upstream node according to the link processing capacity. Using the global time reference phase as the origin of the local clock reference, and combining it with the remaining effective delay budget, the local time slot search window is determined; Within the local time slot search window, the target time slot allocation parameters are obtained by constructing a local conflict graph for the corresponding node and solving the weighted maximum independent set of the local conflict graph.

5. The method according to claim 4, characterized in that, Within the local time slot search window, by constructing a local conflict graph for the corresponding nodes and solving for the weighted maximum independent set in the local conflict graph, the target time slot allocation parameters are obtained, including: Within the local time slot search window, each candidate time slot searched is mapped to a vertex in the local conflict graph; Based on the pre-defined multi-dimensional conflict determination logic, undirected conflict edges are constructed between vertices that have mutual exclusion relationships; The composite semantic weight value corresponding to each vertex is dynamically calculated; the composite semantic weight value is used to characterize the contribution of the selected time slot to the overall system gain and the compatibility impact on subsequent services. Based on the composite semantic weight value, a distributed forward pruning algorithm is executed on the local conflict graph to solve for the weighted maximum independent set. In the weighted maximum independent set, the independent vertex with the largest composite semantic weight value is selected as the target time slot, and the target time slot allocation parameters are determined based on the target time slot.

6. The method according to claim 5, characterized in that, Based on the composite semantic weight values, a distributed forward pruning algorithm is executed on the local conflict graph to solve for the weighted maximum independent set, including: Sort all vertices in descending order of their weight values; The forward pruning process is performed using a greedy heuristic strategy: Select the unprocessed vertex with the highest weight in the current list as the candidate vertex; Check whether the candidate vertex has an edge conflict with the vertices already selected in the set; If there is no conflict, the candidate vertex is added to the temporary optimal time slot set, and the candidate vertex and all its adjacent vertices are removed from the local conflict graph. If a conflict exists, the candidate vertex is discarded and removed from the local conflict graph. Repeat the above pruning process until all vertices in the local conflict graph have been processed, resulting in a locally conflict-free set of independent vertices, which serves as the weighted maximum independent set.

7. The method according to claim 5, characterized in that, The composite semantic weight is calculated by jointly weighting and summing the following four components: Service priority benefit component: The external priority determined based on the service quality identifier of the service flow to be scheduled is jointly determined by the internal priority determined based on the service deadline, transmission period and frame length tuning; Queue delay tolerance component: determined based on the ratio of the difference between the preset maximum end-to-end delay budget of the entire network and the cumulative queuing delay prediction value of the service flow at the current hop node, and used to give weight compensation to urgent service flows; Micro-phase consistency score component: determined based on the reciprocal of the absolute value of the time deviation between the current candidate time slot and the global time reference phase, used to characterize the degree of alignment between time slot selection and the overall network scheduling cycle; The resource fragmentation penalty component is calculated based on the distribution changes of continuous idle intervals of local ports before and after the allocation of candidate time slots.

8. The method according to claim 7, characterized in that, The resource fragmentation penalty component is calculated based on the distribution changes of continuous idle intervals of local ports before and after the allocation of candidate time slots, and includes: Based on the ratio of the length of the continuous free interval in the link port time slot bitmap to the total free capacity, calculate the first resource fragment information entropy before allocation and the second resource fragment information entropy after allocation. The difference between the information entropy of the second resource fragment and the information entropy of the first resource fragment is used as the resource fragment penalty component.

9. A network routing and time slot allocation device across time scales, applied to a converged network including a wireless access domain and a wired domain, characterized in that, include: The first generation unit is used to obtain the global network state characteristics of the converged network according to the scheduling period corresponding to the first time scale, and based on the global network state characteristics, use a preset strategy generation algorithm to generate the target routing path of the service flow to be scheduled in the wired domain and the delay budget constraint parameters and global time reference phase corresponding to the target routing path. The second generation unit is used to obtain the local network status of each hop node of the target routing path according to the scheduling period corresponding to the second time scale, and generate the target time slot allocation parameters of the service flow to be scheduled in each hop node by combining the delay budget constraint parameters and the global time reference phase. The first time scale is in the millisecond range; The second time scale is on the microsecond level; The distribution unit is used to distribute the target time slot allocation parameters determined by each hop node to the gating configuration table of each hop node for locking.

10. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; The memory is used to store computer programs; When the processor executes the program stored in the memory, it implements the method described in any one of claims 1-8.