Hybrid intelligent driven multi-service flow deterministic scheduling method and related device

Abstract

The application is suitable for the technical field of data transmission, and provides a mixed intelligent driving multi-service flow deterministic scheduling method and related equipment, including: calculating a superframe period according to service flow requirements and a clock synchronization signal, and establishing an initialized blank gating table based on the superframe period; modeling network topology information and traffic state of TT flow into a graph structure, using a graph attention network to extract node features of the graph structure, and generating a scheduling action sequence for a hop scheduling granularity through a proximal policy optimization algorithm; encoding the scheduling action sequence into a chromosome of GA, obtaining TT flow time slots for TT flow transmission through a genetic algorithm, updating the initialized blank gating table based on the TT flow time slots, and obtaining an optimal GCL table; in a non-TT flow time slot in the optimal GCL table, determining a priority of an AVB flow according to dynamic characteristics of the AVB flow through an APRA algorithm; and scheduling the AVB flow through a credit value shaper based on an MQSA algorithm according to the priority, and outputting a scheduling sequence of the AVB flow.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of data transmission, and particularly relates to a mixed intelligent driving multi-service flow deterministic scheduling method and related equipment. BACKGROUND

[0002] In low-altitude intelligent networking, the application scenarios such as unmanned aerial vehicle cluster cooperative operation, low-altitude logistics distribution, wide-area inspection and monitoring put forward unprecedentedly strict requirements for data transmission. These scenarios not only require ultra-low latency and high reliability deterministic transmission (such as unmanned aerial vehicle collision avoidance instructions and emergency control signals), but also need to realize the coexistence and efficient transmission of multiple types and multiple priority data flows (key instructions, real-time video, telemetry data, etc.) in a highly dynamic wireless network environment. Time-Sensitive Networking (TSN) technology provides infrastructure support, and its core traffic scheduling mechanism builds a deterministic transmission framework covering time and space resources through fine time slot planning and transmission priority management, and ensures the strict controllability of the performance boundary of critical services.

[0003] The current TSN traffic scheduling technology takes the Time Awareness Shaping (TAS) mechanism as the core framework, and realizes the static allocation of transmission time slots through periodic time slice division. The mainstream scheduling methods are divided into two categories: one is the constraint-based solution space traversal method, such as ILP / SMT, which realizes global optimal solution search through accurate modeling. This method has the advantages of high solving accuracy, but has the disadvantages of low scalability and slow solving speed caused by traversing the entire solution space, and is suitable for small-scale networks. The other is the heuristic algorithm-based scheduling, such as genetic algorithm, tabu search, etc., which quickly generates a feasible solution through iterative optimization. The advantage is fast solving speed, but the solving quality is low, and it is suitable for large-scale networks with low load.

[0004] The existing TSN traffic scheduling algorithm exposes multiple limitations in the low-altitude intelligent networking scenario: the periodic scheduling mechanism based on the Gate Control List (GCL) is difficult to adapt to the dynamic topology of the unmanned aerial vehicle cluster. When the network node experiences link interruption or delay fluctuation due to high mobility, the pre-configured transmission time slot cannot be adaptively adjusted, resulting in impaired deterministic transmission performance of critical control instructions. The traditional ILP / SMT accurate solving method faces the problem of combinatorial explosion when processing large-scale low-altitude networks, far exceeding the millisecond-level scheduling period requirement. Although the heuristic algorithm can speed up the solution, it may cause low solution quality due to convergence deviation when multiple services coexist. In addition, the existing framework does not fully consider the differentiated needs of mixed service flows in the low-altitude scenario. It is difficult to cooperatively guarantee the nanosecond-level synchronization accuracy of safety control instructions and the megabit-level bandwidth demand of video streams under dynamic channel conditions, and it is difficult to meet the needs of high availability and strong determinism of low-altitude intelligent networking. SUMMARY

[0005] Embodiments of the present application provide a hybrid intelligent driving multi-service flow deterministic scheduling method and related equipment, which can solve the problem that in the prior art, when multiple services coexist, the solution quality may be low due to convergence bias, and it is difficult to meet the demand for high availability and strong determinism of low-altitude intelligent networking.

[0006] In a first aspect, embodiments of the present application provide a hybrid intelligent driving multi-service flow deterministic scheduling method, comprising:

[0007] According to the service flow demand and the clock synchronization signal, a superframe period is calculated by a centralized network configuration controller, and an initialized blank gating table is established based on the superframe period;

[0008] The network topology information and the traffic state of the TT flow are modeled as a graph structure, the node features of the graph structure are extracted using a graph attention network, and according to the node features, a scheduling action sequence for a hop scheduling granularity is generated by a proximal policy optimization algorithm;

[0009] The scheduling action sequence is encoded as a chromosome of a GA, TT flow slots for TT flow transmission are obtained by a genetic algorithm, and the initialized blank gating table is updated based on the TT flow slots to obtain a target GCL table;

[0010] According to the dynamic characteristics of the AVB flow, the priority of the AVB flow is determined by an APRA algorithm;

[0011] According to the priority, the AVB flow is scheduled in a non-TT flow slot determined according to the target GCL table by a credit shaper based on an MQSA algorithm, to realize multi-service flow scheduling.

[0012] Optionally, the step of modeling the network topology information and the traffic state of the TT flow as a graph structure, extracting the node features of the graph structure using a graph attention network, and generating a scheduling action sequence for a hop scheduling granularity according to the node features by a proximal policy optimization algorithm, comprises:

[0013] The network topology information and the traffic state of the TT flow are modeled as a time-varying graph structure, a graph attention network with a multi-head bidirectional attention mechanism is used to capture the space-time dependence relationship between flow hop nodes, a five-dimensional state representation covering transmission delay, remaining hop count, scheduling label, port flow transmission time and hop completion time is generated by feature aggregation and gating update as node features;

[0014] According to the node features, a Critic-Actor double-flow decision network is used to generate a scheduling action sequence for a hop scheduling granularity by a PPO algorithm.

[0015] Optionally, before the step of generating a scheduling action sequence for a hop scheduling granularity by a Critic-Actor double-flow decision network using a PPO algorithm according to the node features, it further comprises:

[0016] In the simulation platform, a low-altitude TSN network environment containing dynamic unmanned aerial vehicle nodes, wireless link models and typical service flow scenarios is constructed, and a GAT encoder with a multi-phase double-attention head mechanism, an Actor policy network parameter and a Critic value network are created.

[0017] Based on the current environment state, the GAT encoder generates a high-order state representation according to the five-tuple features of each unmanned aerial vehicle node for the TT flow and the topological edge relationship.

[0018] The old Actor network in the PPO algorithm generates actions by probability sampling, obtains the immediate reward and the next state after the actions interact with the environment, and then generates a complete trajectory, and stores the complete trajectory in an experience replay buffer.

[0019] From the buffer, the trajectory data is collected, the PPO agent loss is calculated through the high-order state representation, the GAT, Actor and Critic network parameters are iteratively updated using gradient descent, and when the indicators such as average round reward, round length and Critic loss change trend stabilize at the expected level, the model is determined to be converged, and a GAT-PPO model for generating a scheduling action sequence for a hop scheduling granularity by a Critic-Actor double-flow decision network using a PPO algorithm is obtained.

[0020] Optionally, the step of encoding the scheduling action sequence into a GA chromosome to obtain TT flow time slots for TT flow transmission and updating the initialized blank gating table based on the TT flow time slots to obtain a target GCL table comprises:

[0021] The scheduling action sequence is encoded into a chromosome matrix, where the row index corresponds to the switching node, the column index corresponds to the time slot, and the gene bit value is the hop ID or the idle identifier.

[0022] The chromosome matrix is verified, and genetic evolution is performed on the verified chromosome matrix to update the global optimal solution counter.

[0023] When there is no improvement for Thre1 consecutive generations, a feasible hop probability distribution is output based on the chromosome state, high-yield hops are inserted into the idle time slots of the initialized blank gating table under the condition of meeting the constraint conditions, and the optimization direction is guided through the reward signal, and after completion, the global optimal solution counter is reset to obtain an elite chromosome.

[0024] The elite chromosome is decoded into a state-action trajectory and stored in an experience replay pool, and the Actor network parameters are fine-tuned using elite samples;

[0025] After reaching the maximum number of iterations, the global optimal chromosome is decoded to generate a final GCL schedule table.

[0026] Optionally, the step of determining the priority of the AVB stream according to the dynamic characteristics of the AVB stream through the APRA algorithm comprises:

[0027] When the AVB stream arrives, the initial expected priority of the AVB stream is calculated based on the deadline of the AVB stream, and a default priority adjustment parameter is set for the AVB stream;

[0028] At the end of each time slot, the current demand rate and the expected completion time of the AVB stream are calculated;

[0029] The utility value of the current time slot is calculated according to the current demand rate and the expected completion time;

[0030] The utility gradient is determined by comparing the changes in the utility values of adjacent time slots, and the priority adjustment parameter is updated according to the utility gradient;

[0031] Based on the updated priority adjustment parameter, the expected priority of the AVB stream is probabilistically mapped to the actual priority.

[0032] Optionally, the step of scheduling the AVB stream in the non-TT stream time slot determined according to the target GCL table to realize multi-service flow scheduling through the credit value shaper based on the MQSA algorithm according to the priority comprises:

[0033] According to the target GCL table, it is determined whether the current time slot is a non-TT stream transmission time slot, and if the current time slot is a non-TT stream transmission time slot, it is determined whether CreditA under the current time slot is greater than zero:

[0034] If CreditA under the current time slot is greater than zero, it is determined whether the current queue is empty, and if it is empty, CreditA is set to zero;

[0035] If the current queue is not empty, it is determined whether the AVB stream with the highest priority can be scheduled, and if it can be scheduled, the AVB stream with the highest priority is allocated the current time slot, and CreditA and CreditB are updated through the credit value shaper during transmission;

[0036] If CreditA under the current time slot is less than or equal to zero, it is determined whether CreditB under the current time slot is greater than zero;

[0037] If CreditB under the current time slot is greater than zero, it is determined whether the current queue is empty, and if it is empty, CreditB is set to zero;

[0038] If the current queue is not zero, it is determined whether the highest priority AVB flow can be scheduled, and if so, the current time slot is allocated to the highest priority AVB flow, and CreditA and CreditB are updated in the transmission process through a credit shaper;

[0039] If CreditB under the current time slot is less than or equal to zero, a time slot is allocated to a BE flow according to the FIFO principle.

[0040] In a second aspect, the embodiments of the present application provide a hybrid intelligent driving multi-service flow deterministic scheduling device, characterized in that the device comprises:

[0041] An initialization module is configured to calculate a superframe period through a centralized network configuration controller according to service flow requirements and a clock synchronization signal, and establish an initialized blank gating table based on the superframe period;

[0042] A construction module is configured to model network topology information and traffic states of TT flows into a graph structure, extract node features of the graph structure using a graph attention network, and generate a scheduling action sequence for an ASR granularity through a proximal policy optimization algorithm according to the node features;

[0043] An encoding module is configured to encode the scheduling action sequence into a chromosome of a GA, obtain TT flow time slots for TT flow transmission through a genetic algorithm, and update the initialized blank gating table based on the TT flow time slots to obtain a target GCL table;

[0044] A priority determination module is configured to determine priorities of AVB flows through an APRA algorithm according to dynamic characteristics of the AVB flows;

[0045] A scheduling module is configured to schedule AVB flows in non-TT flow time slots determined according to the target GCL table through a credit shaper based on an MQSA algorithm according to the priorities, to implement multi-service flow scheduling.

[0046] In a third aspect, the embodiments of the present application provide an electronic device, which comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the hybrid intelligent driving multi-service flow deterministic scheduling method of any one of the above first aspect when executing the computer program.

[0047] In a fourth aspect, the embodiments of the present application provide a computer readable storage medium, which stores a computer program, and the computer program is executable on a processor to implement the hybrid intelligent driving multi-service flow deterministic scheduling method of any one of the above first aspect.

[0048] In a fifth aspect, the embodiments of the present application provide a computer program product, which, when running on a terminal device, causes the terminal device to perform the hybrid intelligent driven multi-service flow deterministic scheduling method of any one of the first aspect.

[0049] It can be understood that the beneficial effects of the second aspect to the fifth aspect described above can be referred to the related description in the first aspect, which will not be repeated here.

[0050] The beneficial effects of the embodiments of the present application compared with the prior art are: a time triggered (TT) flow dynamic scheduling module adopting deep reinforcement learning (DRL) and genetic algorithm (GA) collaborative optimization, a graph attention network (GAT) enhanced proximal policy optimization (PPO) algorithm is used to realize the rapid generation and online readjustment of super cycle GCL; at the same time, an audio and video bridge flow optimization module based on demand rate driven priority mapping and multi-level queue scheduling architecture is constructed. Through the above technical fusion, the method significantly improves the scheduling robustness to cope with dynamic link interruption and topology reconstruction, realizes dynamic balance between solving efficiency and quality, and fully excavates the potential of non-time triggered time slot resources. It can simultaneously meet the ultra-reliable and ultra-low latency transmission requirements of low-altitude safety critical services and the high throughput and low jitter transmission requirements of soft real-time services, and provide adaptive real-time communication core support for urban air traffic, unmanned aerial vehicle cluster cooperation and other scenarios. For the TT flow in the TSN traffic, a time triggered flow online scheduling module based on deep reinforcement learning is proposed. Through the collaborative mechanism of DRL and GA, the scheduling efficiency and robustness are significantly improved. GAT effectively models the network topology and captures the inter-hop dependency, which enhances the algorithm's adaptability to network dynamics; the bidirectional collaborative mechanism of GAT-PPO and GA effectively balances the scheduling solving speed and the quality of the solution, and can quickly generate the optimal GCL scheduling table in a complex dynamic network environment. For the AVB flow in the TSN traffic, an audio and video bridge flow online scheduling module based on adaptive priority is proposed. Based on adaptive priority adjustment and queue scheduling, the idle resource window reserved by the TT flow scheduling is fully utilized. The APRA algorithm dynamically calculates and maps the priority according to the deadline, data volume and real-time network state of the flow; the MQSA algorithm combines the priority with the CBS mechanism to dynamically determine the transmission order of the AVB flow. This mechanism significantly increases the transmittable opportunity of the AVB flow and effectively reduces its transmission delay and jitter without affecting the scheduled TT flow. BRIEF DESCRIPTION OF DRAWINGS

[0051] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description only constitute some embodiments of the present application, and other drawings can be obtained by those skilled in the art without any creative effort.

[0052] Figure 1 is a flow chart of a hybrid intelligent driving multi-service flow deterministic scheduling method provided by the embodiments of the present application;

[0053] Figure 2 is a running schematic diagram of a hybrid intelligent driving multi-service flow deterministic scheduling method provided by the embodiments of the present application;

[0054] Figure 3 is a flow chart of a GRAD-GA algorithm of a hybrid intelligent driving multi-service flow deterministic scheduling method provided by the embodiments of the present application;

[0055] Figure 4 is a flow chart of a MQSA algorithm of a hybrid intelligent driving multi-service flow deterministic scheduling method provided by the embodiments of the present application;

[0056] Figure 5 is a structural schematic diagram of a hybrid intelligent driving multi-service flow deterministic scheduling device provided by the embodiments of the present application;

[0057] Figure 6 is a structural schematic diagram of an electronic device provided by the embodiments of the present application;

[0058] Figure 7 is a structural schematic diagram of a computer readable storage medium provided by the embodiments of the present application. DETAILED DESCRIPTION

[0059] In the following description, specific details are set forth in order to provide a thorough understanding of the embodiments of the present application. However, persons skilled in the art will understand that the present application can be practiced without these specific details. In other instances, well-known systems, structures, circuits, and methods have not been described in detail in order to avoid obscuring the description of the present application.

[0060] It should be understood that when used in the specification and the appended claims of the present application, the term "comprising" indicates the presence of the described features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0061] It should also be understood that the term “and / or” as used herein, means any one or more of the associated listed items, and that includes one or more of the associated listed items, as well as all possible combinations of the associated listed items.

[0062] As used in the description of the application and the appended claims, the term “if’ can be interpreted to mean “when” or “upon” or “in response to determining” or “in response to detecting” depending on the context. Similarly, the phrase “if it is determined” or “if [a described condition or event] is detected” can be interpreted to mean “upon determining” or “in response to determining” or “upon [the described condition or event] being detected” or “in response to [the described condition or event] being detected,” depending on the context.

[0063] In addition, in the description and the appended claims of the application, the terms “first”, “second”, “third”, etc. are used only to distinguish descriptions, and cannot be understood as indicating or implying relative importance.

[0064] Reference in the specification to “one embodiment” or “some embodiments” means that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment of the application. The appearances of the phrase “in one embodiment” or “in some embodiments” in various places in the specification are not necessarily all referring to the same embodiment, although it can. The terms “including,” “containing,” “having,” and variations thereof are meant to encompass the terms “including but not limited to,” unless otherwise indicated.

[0065] As Figure 1 As shown in the figure, the embodiments of the application provide a hybrid intelligent driving multi-service flow deterministic scheduling method, which comprises:

[0066] S101, according to the service flow demand and the clock synchronization signal, calculating the superframe period by the centralized network configuration controller, and establishing the initialized blank gating table based on the superframe period;

[0067] S102, modeling the network topology information and the traffic state of the TT flow as a graph structure, extracting the node features of the graph structure using the graph attention network, and generating the scheduling action sequence for the hop scheduling granularity according to the node features through the proximal policy optimization algorithm;

[0068] S103, encoding the scheduling action sequence into the chromosome of GA, obtaining the TT flow time slot of the TT flow transmission through the genetic algorithm, and updating the initialized blank gating table based on the TT flow time slot to obtain the target GCL table.

[0069] S104, determining the priority of the AVB stream according to the dynamic characteristics of the AVB stream through an adaptive priority reordering algorithm (APRA);

[0070] S105, scheduling the AVB stream in the non-TT stream time slot determined according to the target GCL table based on a multi-level queue scheduling algorithm (MQSA) through a credit shaper according to the priority, to realize multi-service flow scheduling.

[0071] Among them, the TT flow is a time triggered flow, the AVB flow is an audio and video data flow transmitted based on the AVB (Audio Video Bridging) technology, and the BE flow is a background load mixed with the time triggered (TT) flow for transmission, which is used to verify the time determinacy of the TT flow under high load.

[0072] As shown in the example, Figure 2 The embodiment of the application provides a mixed intelligent driving multi-service flow determinacy scheduling method, and the steps are as follows:

[0073] The unmanned aerial vehicle and the gateway node send registration information (including position, capability matrix) to a centralized user configuration (CUC), and the CUC stores the registration information into a database after checking airspace compliance. The whole process uses ACK confirmation information to ensure success.

[0074] A centralized network configuration (CNC) periodically detects network topology information to obtain the positions and link quality between nodes. After fusing meteorological radar and airspace control data, the database is updated.

[0075] The CNC sends a synchronization signal to all nodes based on a GNSS / PTP hybrid clock synchronization protocol to establish a master-slave clock tree.

[0076] Each node sends the service flow information that needs to be sent to the CUC, and the CUC stores the registration information into the database after registration, and forwards the information to the CNC through a QoS mapping protocol.

[0077] The CNC obtains all real-time flow information, including a TT flow set , an AVB flow set , and a BE flow set . According to the received TSN flow information, the superframe period mT is calculated, and the blank GCL scheduling table is initialized.

[0078] The CNC calls the TT flow scheduling module to allocate GCL time slots to the TT flow through the GAT-PPO model trained offline combined with the genetic algorithm, thereby reserving resources for the TT flow. The GAT-PPO model is continuously improved by using the feedback of the genetic algorithm.

[0079] The CNC sends the scheduling configuration result to each TSN node through a remote management protocol.

[0080] Each terminal node starts to send service data flow to the TSN node, and the TSN node shapes and allocates the data flow to different queues according to the scheduling result.

[0081] Each node schedules the AVB flow through the AVB flow online scheduling module in the non-TT flow transmission time slot. The non-TT flow is also considered in the same superframe period mT, and the BE flow is scheduled and allocated according to the FIFO algorithm (first-in-first-out algorithm) in the non-TT flow transmission time.

[0082] The CNC periodically probes and checks the transmission of each node, the link state, and the burst flow, and makes online local adjustments to the GCL table through the TT flow online scheduling module.

[0083] In one possible implementation, the step of modeling the network topology information and the traffic state of the TT flow as a graph structure, using a graph attention network to extract node features of the graph structure, and generating a scheduling action sequence for a hop scheduling granularity according to the node features through a proximal policy optimization algorithm, comprises:

[0084] The network topology information and the traffic state of the TT flow are modeled as a time-varying graph structure, a graph attention network with a multi-head bidirectional attention mechanism is used to capture the space-time dependency relationship between flow hop nodes, and a five-dimensional state representation covering transmission delay, remaining hop count, scheduling label, port flow sending time, and hop completion time is generated through feature aggregation and gating update as node features;

[0085] According to the node features, a Critic-Actor double-flow decision network is used to generate a scheduling action sequence for a hop scheduling granularity through a PPO algorithm.

[0086] For example, the network topology and the traffic state are modeled as a time-varying graph structure, a multi-head bidirectional attention mechanism is used to capture the space-time dependency relationship between flow hop nodes, and a high-order state representation that integrates the topology context is generated through feature aggregation and gating update; then a Critic-Actor double-flow decision network is constructed, a PPO algorithm is used to map state information such as flow hop level queue dynamics and link delay to a diverse scheduling sequence. The specific process is described as follows:

[0087] Using the shortest path first algorithm, the transmission path for each TT stream can be obtained, that is, for each TT stream... All of these can yield a set of hop points, such as The transmission path is ,but The set of hops is Each hop in the entire TT flow is modeled as a node of a GNN. To overcome the limitations of traditional flow-level scheduling models, the five-tuple state characteristics for hop-point scheduling granularity are defined as follows:

[0088] (1)

[0089] in, Indicates the transmission delay of the current hop. This represents the remaining hops for the corresponding flow. It is a scheduling label, if If it is true, it means that the jump has been scheduled; otherwise, if If the jump is not scheduled, it means that the jump has not been scheduled. This indicates when the port can send a new stream. This indicates that the jump is in a certain state. The earliest completion time, disregarding queuing delays, is the actual completion time if the hop has been scheduled; otherwise, it is the sum of the actual completion time of the previous hop and the transmission delay of this hop. This quintuple design overcomes the shortcomings of a single perspective by considering the global progress of the flow (…). ) and the real-time resource status of key nodes ( The efficient coupling of heterogeneous information such as network topology evolution trends and micro-resource bottlenecks enables the DRL strategy to simultaneously perceive network topology evolution trends and micro-resource bottlenecks, which is the core foundation for supporting highly robust "skip-level" scheduling oriented towards hop-point scheduling granularity.

[0090] The dependencies between nodes are modeled as directed edges in a graph, such as edges... If it exists, it means The corresponding skip scheduling must be in the scheduling The corresponding jump is completed beforehand, and the set of edges is denoted as... Each node then has its set of predecessors. and the set of successor jumps .

[0091] set up Each attention head, in each iteration, comes from... Half of the attention heads are allocated to collecting topological information from the predecessor hop set, while the remaining half is used to collect topological information from the successor hop set. The normalized attention coefficients are defined as follows:

[0092] (2)

[0093] Iteration Next time Relative to attention head z The relative importance of. Among them, For the trainable vectors of the iterative process, Let be the trainable weight matrix during the iteration process. The aggregation function based on multi-head attention is calculated as follows:

[0094] (3)

[0095] Then, by applying the Exponential Linear Unit (ELU) activation in the Update function, we can obtain:

[0096] (4)

[0097] The Aggregate function is used to accumulate topological information about neighbors. Then, a non-linear Update function is used to modify the accumulated information. This process is repeated L times to obtain the characteristics of each hop.

[0098] Next, the TT flow scheduling is expressed as a Markov Decision Process (MDP) with state, action, and reward representations. The state is defined as a quintuple (including transmission delay, remaining hops, etc.) consistent with the features of the GNN nodes. The feature vector generated after processing the original features through the GNN's multi-head bidirectional attention mechanism serves as the input state for the reinforcement learning agent, enabling the policy network to make scheduling decisions based on information from the fused topological context.

[0099] action The scheduling is defined as each node selecting one hop from the set of feasible hops in the current time slot for scheduling. To reduce the algorithm complexity for each time slot selection and accelerate the convergence speed in the early stages of the algorithm, four greedy heuristics are proposed to assist the selection action while ensuring that the flow arrives on time:

[0100] Prioritize transmission delay: prioritize forwarding traffic with the shortest transmission delay.

[0101] Remaining hops priority: Prioritize forwarding traffic with the most remaining hops.

[0102] Scheduled time priority: Prioritize traffic with the shortest scheduled time (the maximum of the predecessor hop completion time and port idle time) for forwarding.

[0103] Earliest completion time priority: Prioritize forwarding traffic with the earliest completion time.

[0104] A multi-heuristic polling mechanism and an ε-perturbation strategy are employed to ensure diversity: four heuristic strategies are activated alternately in a random order, while an ε-perturbation mechanism is introduced to randomly select any jump from the feasible jump set with probability ε (rather than following the current heuristic). This randomness breaks the deterministic limitations of heuristic strategies. Compared to a single ε-greedy approach, this method can accelerate convergence by leveraging domain knowledge while ensuring sufficient exploration through polling and random perturbation, avoiding premature entrapment in local optima dominated by a single heuristic rule.

[0105] award The fitness increment is defined as the step size in the genetic algorithm. The reward function is defined as follows:

[0106] (5)

[0107] in, Indicates selecting an action The fitness value of the individual. This indicates the selection of an action. The fitness value of the previous individual.

[0108] After each state transition, a check is required. Does the indicated scheduling violate the constraints? The constraints are as follows:

[0109] In a time slot, each stream can only be transmitted on one port, and each port can only process one stream.

[0110] The end-to-end delay of a stream must be less than its due date.

[0111] Within a supercycle, some flows may be scheduled multiple times. To ensure no interference between repeated scheduling cycles, the first... The transmission and reception times of the stream within each cycle must adhere to the following constraints:

[0112] (6)

[0113] in, Indicates the time the stream was sent. Indicates the time of stream reception. It is a supercycle.

[0114] Because frame buffer resources are limited, streams must meet the following constraints when queuing:

[0115] (7)

[0116] (8)

[0117] (9)

[0118] in, and Indicates the first The used and idle queues of each scheduling node. Indicates the queue length.

[0119] Based on the network input generated by GNN, the Actor network will process each... Generate a scalar Then through Function-generated action probability distribution:

[0120] (10)

[0121] The PPO algorithm contains two Actor networks, one of which is called the old Actor network. The parameters are Another one is called the Actor network. The parameters are . Used to interact with the environment to obtain a trajectory, Use these trajectories for K parameter updates. The loss function of the Actor network is a pruning agent objective function, defined as follows:

[0122] (11)

[0123] in, This represents the ratio between the probabilities of the outputs of two Actor networks. For advantageous space, This represents the pruning operation, used to measure the difference between two networks, and is calculated as follows:

[0124] (12)

[0125] Based on time difference theory, the value loss function of the Critic network is defined as follows:

[0126] (13)

[0127] in, The state calculated by the Critic network based on the GNN input. The estimated value. The calculation method is as follows:

[0128] (14)

[0129] in, Let be the discount factor. Finally, the loss function for PPO is derived, defined as follows:

[0130] (15)

[0131] in, and There are two weighting coefficients. It is the entropy reward function, calculated as follows:

[0132] (16)

[0133] In one possible implementation, before the step of generating a sequence of scheduling actions at the hop-point scheduling granularity using the PPO algorithm through a Critic-Actor dual-stream decision network based on the node characteristics, the method further includes:

[0134] In the simulation platform, a low-altitude TSN network environment including dynamic UAV nodes, wireless link models and typical service flow scenarios is constructed, and a GAT encoder with a multi-phase dual attention head mechanism, Actor policy network parameters and Critic value network are created.

[0135] Based on the current environmental state, the GAT encoder is used to generate a high-order state representation according to the five-tuple features and topological edge relationships of each UAV node for the TT flow.

[0136] The old Actor network in the PPO algorithm generates actions through probability sampling, obtains immediate rewards and the next state after the actions interact with the environment, and then generates a complete trajectory, which is then stored in the experience replay buffer.

[0137] Trajectory data is collected from the buffer, and the PPO agent loss is calculated through the higher-order state representation. Gradient descent is used to iteratively update the parameters of the GAT, Actor, and Critic networks until a preset number of training rounds are completed. The changing trends of average round reward, round length, and Critic loss are used as indicators. When the indicators stabilize at the expected level, the model is determined to have converged, resulting in the GAT-PPO model used to generate scheduling action sequences oriented towards hop scheduling granularity through the Critic-Actor dual-stream decision network and the PPO algorithm.

[0138] For example, in summary, the overall steps for offline training of the GAT-PPO model are as follows:

[0139] Building the training environment and initializing the model: First, a low-altitude TSN network environment including dynamic UAV nodes, wireless link models, and typical service flow scenarios is built in the simulation platform; then, the core components are initialized: a multi-phase dual-attention head mechanism GAT encoder, Actor policy network parameters, and Critic value network are created, and their weight parameters are set, and the training epochs are determined. And the number of parameter updates K in each round.

[0140] Start Training Rounds and Policy Freeze: At the beginning of each training round, reset the simulation environment to its initial state and freeze the old Actor network parameters. Synchronize with the current Actor parameters ,freeze Used for sampling all interactions within this round.

[0141] Interactive sampling generates empirical data: Based on the current environment state, the GAT encoder is used to process the node quintuple features and topological edge relationships to generate higher-order state representations. The old Actor network generates actions through probability sampling (sampling actions according to the network output distribution with a probability of 1-ε, or randomly activating one of four heuristics with a probability of ε). Get instant rewards after interacting with the environment. and the next state Complete trajectory Store in the experience replay buffer.

[0142] Strategy optimization and model update: Extract node features from the sampled trajectory data in the buffer using the GAT encoder. The PPO proxy loss is calculated, and gradient descent is used to update the parameters of the GAT, Actor, and Critic networks K times.

[0143] Convergence evaluation and model export: Completed After each training round, monitor the trends of average round reward, round length, and Critic loss. Model convergence is determined when these metrics stabilize at expected levels. Save the final training output: the optimized Actor policy network weights. Critic value network weights It serves as the decision engine for the subsequent online scheduling module.

[0144] In one possible implementation, the steps of encoding the scheduling action sequence as a chromosome of a GA, obtaining the TT stream time slots for TT stream transmission through a genetic algorithm, and updating the initially blank gating table based on the TT stream time slots to obtain the target GCL table include:

[0145] The scheduling action sequence is encoded into a chromosome matrix, where the row index corresponds to the exchange node, the column index corresponds to the time slot, and the gene bit value is either the jump ID or the idle identifier.

[0146] The chromosome matrix is ​​validated, and genetic evolution is performed on the validated chromosome matrix to update the global optimal solution counter;

[0147] When no improvement is achieved in the first Thre1 generation, the feasible jump probability distribution is output based on the chromosome state. Under the condition of satisfying the constraints, a high-yield jump is inserted into the idle time slot of the initialized blank gating table, and the optimization direction is guided by the reward signal. After completion, the global optimal solution counter is reset to obtain the elite chromosome.

[0148] The elite chromosomes are decoded into state-action trajectories and stored in the experience replay pool, and the Actor network parameters are fine-tuned using elite samples.

[0149] After reaching the maximum number of iterations, the globally optimal chromosome is decoded to generate the final GCL scheduling table.

[0150] For example, such as Figure 3 As shown, each individual can calculate a fitness value to measure its quality. The fitness value is related to the end-to-end delay and is defined as follows:

[0151] (17)

[0152] in, For flow The last hop, end-to-end delay, For flow In state The completion time of the last hop is the actual completion time if it has been scheduled; otherwise, it is the sum of the actual completion time of the previous hop and the transmission delay of this hop.

[0153] During the population initialization phase, the GAT-PPO model drives the Actor network to sample action sequences from the feasible hop set through an ε-greedy strategy. A certain proportion (preferably 30%~50%) of actions are sampled using a random heuristic round-robin approach (alternating between strategies such as transmission delay priority and remaining hop priority), while the remaining proportion of actions are sampled probabilistically using a policy network to generate diverse action sequences. This sequence is a time-ordered set of hop scheduling decisions, with each decision unit defined as a binary tuple: action = (slot index, hop ID). The action sequence is then converted into a GA chromosome—encoding hop IDs in integer matrix form. The chromosome is segmented according to the exchange node index, with each segment corresponding to the maximum number of slots. The gene position (i,j) represents the scheduling decision of the i-th exchange node in the j-th slot: non-zero values ​​represent hop IDs (e.g., 1, 2, 4), and 0 indicates no scheduling. Constraint checks are performed on the generated chromosome, eliminating individuals that violate the constraints, and resampling continues until a valid solution is generated.

[0154] repeat Second-rate( (where GA is the population size), forming an initial population containing different scheduling strategies, where each individual corresponds to a feasible scheduling scheme that integrates topology awareness and latency optimization.

[0155] During the population evolution phase, the algorithm performs operations on each of the Np individuals in each generation from generation 1 to the maximum generation Gen: new individuals are generated through bi-point crossover with probability α (typically 0.6~0.9) to enhance population diversity; gene loci are exchanged and mutated with probability β (typically 0.01~0.1) to escape local optima; then, the fitness values ​​of all individuals are calculated. If the local optimum is better than the global optimum, the global optimum is updated and count1 is incremented by 1; otherwise, count2 is incremented by 1. When count1 reaches the threshold Thre1 (preferred empirical value is 6~10), the GAT-PPO-assisted local search mechanism is activated: the Actor network outputs the probability distribution of feasible jumps based on the current chromosome state (such as the time window of unscheduled tasks, resource consumption, etc.), randomly selects idle time slots for insertion under the premise of satisfying constraints such as port exclusivity and time windows, evaluates the action benefit by calculating the reward value, and guides the feasible jump insertion strategy to optimize in the direction of improving fitness. After completion, count1 is reset to recount the continuous improvement of generations. If count2 is less than the threshold Thre2 (usually 3 times Thre1), then an elite strategy is used to retain the best individual in the current population to accelerate convergence; at the same time, a probability-based strategy is employed. The algorithm employs an individual update strategy, randomly introducing new individuals to prevent stagnation until the Gen iterations are completed, at which point the final optimal solution is output. Simultaneously, the genetic algorithm decodes each generation of elites into state-action trajectories, storing them in the GAT-PPO experience replay pool. This triggers asynchronous fine-tuning of the policy network—the PPO algorithm uses elite samples to update Actor parameters, reinforcing the high-reward scheduling mode.

[0156] Specifically, the steps of encoding the scheduling action sequence into a chromosome of a GA, obtaining the TT stream time slots for TT stream transmission through a genetic algorithm, and updating the initial blank gating table based on the TT stream time slots to obtain the target GCL table include:

[0157] Population initialization: GAT-PPO generates scheduling jump sequence through ε-greedy strategy (30% random heuristic polling + 70% strategy network sampling), and encodes it into chromosome matrix: row index corresponds to exchange node, column index corresponds to time slot, and gene bit value is jump ID (non-zero) or idle identifier (0).

[0158] Constraint verification: Check the port exclusivity, time window and overcycle interference constraints of the initial chromosome, remove illegal solutions and resample to ensure that every individual in the population is a feasible scheduling scheme.

[0159] Genetic evolution: Two-point crossover is performed with probability α to enhance diversity, and gene loci are exchanged with probability β to achieve mutation; the fitness value of an individual (total end-to-end delay of TT flow) is calculated, and the global optimal solution counter (count1 / count2) is updated.

[0160] Local optimization: When no improvement is achieved in the previous Thre1 generations, a local search is initiated: The Actor network outputs the probability distribution of feasible jumps based on the chromosome state, inserts high-yield jumps into idle time slots under the constraints, and guides the optimization direction through the reward signal (fitness increment). After completion, count1 is reset.

[0161] Strategy Feedback: Each generation of elite chromosomes is decoded into state-action trajectories and stored in the experience replay pool, triggering asynchronous policy updates of PPO—using elite samples to fine-tune the Actor network parameters and enhance the high-reward scheduling mode.

[0162] Output termination: After reaching the maximum number of iterations (Gen), decode the globally optimal chromosome: map the gene position (i,j) to the scheduling hop ID of the j-th time slot of the i-th exchange node (0 indicates idle), and generate the final GCL scheduling table.

[0163] In one possible implementation, the step of determining the priority of an AVB stream using the APRA algorithm based on the dynamic characteristics of the AVB stream includes:

[0164] When an AVB stream arrives, the initial expected priority of the AVB stream is calculated based on the AVB stream's deadline, and a default priority adjustment parameter is set for the AVB stream.

[0165] At the end of each time slot, calculate the current demand rate and estimated completion time of the AVB stream;

[0166] Calculate the utility value of the current time slot based on the current demand rate and the estimated completion time;

[0167] Compare the changes in utility values ​​between adjacent time slots to determine the utility gradient, and update the priority adjustment parameters based on the utility gradient;

[0168] Based on the updated priority adjustment parameters, the expected priority of the AVB stream is probabilistically mapped to the actual priority.

[0169] For example, AVB streams can be categorized based on their original priority. flow and harmony Stream. This module relies on priority adjustment parameters. Flows of equal priority are further dynamically assigned priorities with hop-point scheduling granularity. When a flow arrives at a switching node, the switching node will, according to... The expected priority of a flow is dynamically adjusted, mapping it to its actual priority. The expected priority of a flow is calculated based on its deadline; the shorter the deadline, the higher the priority. The core parameters of this mapping process are... The calculation process is as follows.

[0170] A higher ratio of bytes to deadline indicates a higher minimum transmission rate required by the system. When actual network conditions (such as bandwidth and congestion) cannot meet this rate, the stream completion time is more likely to exceed the latency requirement, leading to an increased deadline miss rate. Therefore, this paper defines this ratio as the demand rate, as shown in the following formula:

[0171] (18)

[0172] in, Indicates the first Number of bytes per AVB stream Indicates the first The deadline for each AVB stream. This invention innovatively proposes a utility function for evaluating the scheduling urgency of AVB streams. At the end of each time slot, the switching node calculates the stream according to the following formula. The utility value,

[0173] (19)

[0174] in, , It is a constant and satisfies . Let the expected flow completion time be denoted as under the current network conditions, calculated using network calculus theory. This utility function captures the flow's demand rate to measure its absolute importance, and captures the flow's actual performance and the current network state to measure its relative importance. It is strictly convex, meaning that the local optimum is the globally unique optimum. Next, by continuously adjusting the priority configuration according to the gradient direction, we can gradually approach and ultimately maximize the network utility.

[0175] In the At the end of the first time slot, the switching node calculates the first time slot according to equation (20). and the The utility value of each time slot is respectively used as and This indicates that the first... The utility gradient for each time slot is:

[0176] (20)

[0177] in, and The flow in the first place is respectively and the The demand rate for each time slot. Then the utility gradient. Transform into The step size of the change is calculated as follows:

[0178] (twenty one)

[0179] in, It is the conversion factor, set to a conservative small value. and They represent The upper and lower bounds satisfy To prevent due to Too low a level can lead to starvation or An excessively high priority will cause a large number of streams to be assigned a higher priority, thus affecting the performance of other high-priority streams.

[0180] For each flow, the switching node will be based on Probabilistically prioritize the expected flow Mapping to actual priority Specifically:

[0181] if Then the swapping node uses The expected priority of the flow is assigned to the flow based on the probability; or... The probability will Mapped to the next lower priority, i.e. ;if If it's already the lowest priority, then set... .

[0182] if Then the swapping node uses The expected priority of the flow is assigned to the flow based on the probability; or... The probability will Mapped to the next higher priority, i.e. ;if If it's already the highest priority, then set... .

[0183] Specifically, the step of determining the priority of an AVB stream using the APRA algorithm based on the dynamic characteristics of the AVB stream includes:

[0184] Stream arrival and initialization: When an AVB stream arrives, its initial expected priority is calculated based on the deadline, and the default priority adjustment parameters are set.

[0185] Periodic status assessment: At the end of each time slot, the utility value is assessed according to Equation (20) based on the current demand rate of the computing stream of the exchange node and the estimated completion time derived from the network calculation.

[0186] Gradient-driven parameter update: Compare the changes in utility values ​​between adjacent time slots, calculate the utility gradient according to equation (21), and update the parameters according to equation (22). .

[0187] Probabilistic priority mapping: Mapping expected priorities to actual priorities according to mapping rules.

[0188] Closed-loop scheduling and refresh: The mapped actual priority is applied to queue scheduling, and the deadline is refreshed in the next time slot according to the streaming progress.

[0189] In one possible implementation, the step of scheduling AVB flows according to the priority, based on the MQSA algorithm and using a credit value shaper, within the non-TT flow time slots determined according to the target GCL table, to achieve multi-service flow scheduling includes:

[0190] Based on the target GCL table, determine whether the current time slot is a non-TT stream transmission time slot. If the current time slot is a non-TT stream transmission time slot, determine whether CreditA under the current time slot is greater than zero.

[0191] If CreditA in the current time slot is greater than zero, then check if the current queue is empty. If it is empty, then set CreditA to zero.

[0192] If the current queue is not empty, determine whether the highest priority AVB stream can be scheduled. If it can be scheduled, allocate the current time slot to the highest priority AVB stream and update CreditA and CreditB during transmission using the credit value shaper.

[0193] If CreditA in the current time slot is less than or equal to zero, then determine whether CreditB in the current time slot is greater than zero.

[0194] If CreditB in the current time slot is greater than zero, then check if the current queue is empty. If it is empty, then set CreditB to zero.

[0195] If the current queue is not zero, determine whether the highest priority AVB stream can be scheduled. If it can be scheduled, allocate the current time slot to the highest priority AVB stream and update CreditA and CreditB during transmission through the credit value shaper.

[0196] If CreditB in the current time slot is less than or equal to zero, then a time slot is allocated to the BE stream according to the FIFO principle.

[0197] For example, an AVB stream is a non-TT stream and needs to be transmitted under the control of the GCL. This algorithm is mainly used to solve the conflict problem of multiple AVB streams. It schedules and selects AVB streams in a queue. When there are streams waiting to be transmitted in the queue and they are in an idle state, the next data stream to be transmitted can be selected. The specific steps are as follows: Figure 4 As shown.

[0198] When Class A queue is in the transmission state, its credit value decreases at the idleSendA rate. If it enters the idle state, it immediately checks the stream to be transmitted: if there is data, it selects the stream for transmission using the APRA algorithm and accumulates the Class B credit value at the idleSlopeB rate; if there is no data, it resets CreditA to zero and switches to the Class B state for judgment—if the latter is in the transmission state, its credit value decreases at the idleSendB rate; in the waiting state, it executes the APRA stream selection algorithm and restores the idleSlopeA credit value accumulation of Class A; if Class B also has no stream to be transmitted, it resets CreditB.

[0199] The overall steps of the MQSA algorithm are as follows:

[0200] Each time-slot switching node calculates the priority of all unscheduled AVB streams using the APRA algorithm. Based on the priority, the AVB streams in the queue are reordered, and newly arriving AVB streams are added to their respective queues in sequence.

[0201] Determine whether the GCL is a non-TT stream transmission slot at this time. If not, transmit the TT stream; otherwise, proceed to the next step.

[0202] Determine the current state of CreditA:

[0203] If the value is greater than zero, check if the current queue is empty. If it is empty, set CreditA to zero. Otherwise, check if the highest priority data stream can be scheduled. If it can be scheduled, allocate a time slot for it and continuously decrease CreditA and increase CreditB during transmission. If it cannot be scheduled, proceed to the next step.

[0204] If it is less than or equal to zero, proceed to the next step.

[0205] Determine the current CreditB:

[0206] If the value is greater than zero, check if the current queue is empty. If it is empty, set CreditB to zero. Otherwise, check if the highest priority data stream can be scheduled. If it can be scheduled, allocate a time slot for it and continuously decrease CreditB and increase CreditA during transmission. If it cannot be scheduled, proceed to the next step.

[0207] If it is less than or equal to zero, then a time slot will be allocated to the BE stream according to the FIFO principle.

[0208] The solution is iterated until there are no more data streams in the schedulable interval in the transmission queue.

[0209] In one possible implementation, this application provides a hybrid intelligent-driven multi-service flow deterministic scheduling apparatus, characterized in that the apparatus comprises:

[0210] The initialization module 201 is used to calculate the superframe period through the centralized network configuration controller according to the service flow requirements and clock synchronization signal, and to establish an initial blank gate table based on the superframe period.

[0211] The construction module 202 is used to model the network topology information and the traffic status of TT flow as a graph structure, use a graph attention network to extract the node features of the graph structure, and generate a scheduling action sequence oriented towards hop scheduling granularity based on the node features through a near-end policy optimization algorithm.

[0212] The encoding module 203 is used to encode the scheduling action sequence into a chromosome of GA, obtain the TT stream time slot of TT stream transmission through a genetic algorithm, and update the initial blank gating table based on the TT stream time slot to obtain the target GCL table.

[0213] Priority determination module 204 is used to determine the priority of AVB streams based on their dynamic characteristics using the APRA algorithm.

[0214] The scheduling module 205 is used to schedule AVB streams according to the priority, based on the MQSA algorithm and a credit value shaper, in the non-TT stream time slots determined according to the target GCL table, thereby realizing multi-service stream scheduling.

[0215] In one possible implementation, such as Figure 6As shown, this application embodiment provides a terminal device 300, including: a memory 310, a processor 320, and a computer program 311 stored in the memory 310 and executable on the processor 320. When the processor 320 executes the computer program 311, it performs the following: calculates the superframe period through a centralized network configuration controller according to service flow requirements and clock synchronization signals, and establishes an initial blank gating table based on the superframe period; models the network topology information and the traffic status of TT flow as a graph structure, extracts the node features of the graph structure using a graph attention network, and performs the following based on the... Based on node characteristics, a scheduling action sequence oriented towards hop-point scheduling granularity is generated through a near-end policy optimization algorithm. The scheduling action sequence is encoded as a chromosome of the GA. The TT stream time slots for TT stream transmission are obtained through a genetic algorithm, and the initial blank gating table is updated based on the TT stream time slots to obtain the target GCL table. According to the dynamic characteristics of AVB streams, the priority of AVB streams is determined through the APRA algorithm. According to the priority, the AVB streams are scheduled in the non-TT stream time slots determined by the target GCL table through the credit value shaper based on the MQSA algorithm, thereby realizing multi-service stream scheduling.

[0216] In one possible implementation, such as Figure 7 As shown, this application embodiment provides a computer-readable storage medium 400, on which a computer program 411 is stored. When executed by a processor, the computer program 411 performs the following: Calculates the superframe period using a centralized network configuration controller based on service flow requirements and clock synchronization signals, and establishes an initial blank gating table based on the superframe period; models the network topology information and the traffic status of TT flows as a graph structure, extracts the node features of the graph structure using a graph attention network, and generates a scheduling action sequence oriented towards hop-point scheduling granularity using a near-end policy optimization algorithm based on the node features; encodes the scheduling action sequence as chromosomes of a GA, obtains the TT flow time slots for TT flow transmission using a genetic algorithm, and updates the initial blank gating table based on the TT flow time slots to obtain a target GCL table; determines the priority of AVB flows using the APRA algorithm based on the dynamic characteristics of AVB flows; and schedules AVB flows according to the priority using a credit value shaper based on the MQSA algorithm in the non-TT flow time slots determined according to the target GCL table, thereby realizing multi-service flow scheduling.

[0217] It should be noted that the information interaction and execution process between the above-mentioned devices / units are based on the same concept as the method embodiments of this application. For details on their specific functions and technical effects, please refer to the method embodiments section, and they will not be repeated here.

[0218] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments 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. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0219] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of this application can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include at least: any entity or device capable of carrying computer program code to a photographing device / terminal device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium. Examples include USB flash drives, portable hard drives, magnetic disks, or optical disks. In some jurisdictions, according to legislation and patent practice, computer-readable media cannot be electrical carrier signals or telecommunication signals.

[0220] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0221] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0222] In the embodiments provided in this application, it should be understood that the disclosed apparatus / network devices and methods can be implemented in other ways. For example, the apparatus / network device embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0223] 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.

[0224] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications 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 this application, and should all be included within the protection scope of this application.

Claims

1. A hybrid intelligent driven multi-traffic flow deterministic scheduling method, characterized in that, The method comprises the following steps: According to the business flow requirements and the clock synchronization signal, the centralized network configuration controller calculates the superframe period, and establishes an initialized blank gating table based on the superframe period; The network topology information and the traffic state of the TT flow are modeled as a graph structure, the node features of the graph structure are extracted using a graph attention network, and a scheduling action sequence for a hop scheduling granularity is generated based on the node features through a proximal policy optimization algorithm; The step of modeling the network topology information and the traffic state of the TT flow as a graph structure, extracting the node features of the graph structure using a graph attention network, and generating a scheduling action sequence for a hop scheduling granularity based on the node features through a proximal policy optimization algorithm comprises: The network topology information and the traffic state of the TT flow are modeled as a time-varying graph structure, a graph attention network with a multi-head bidirectional attention mechanism is used to capture the spatio-temporal dependence relationship between flow hop nodes, and a five-dimensional state representation covering transmission delay, remaining hop count, scheduling label, port flow transmission time and hop completion time is generated through feature aggregation and gating update as node features; A low-altitude TSN network environment containing dynamic unmanned aerial vehicle nodes, wireless link models and typical business flow scenarios is constructed in a simulation platform, and a GAT encoder with a multi-phase double attention head mechanism, Actor policy network parameters and Critic value network are created; Based on the current environment state, the GAT encoder generates high-order state representations according to the five-tuple features of the TT flow of each unmanned aerial vehicle node and the topological edge relationship; The old Actor network in the PPO algorithm generates actions through probability sampling, and obtains immediate rewards and the next state after the actions interact with the environment, thereby generating complete trajectories, and storing the complete trajectories in an experience replay buffer; Trajectory data is collected from the buffer, the PPO agent loss is calculated through the high-order state representations, the GAT, Actor and Critic network parameters are iteratively updated using gradient descent, and the training is completed after a predetermined number of rounds, and the average round reward, round length and Critic loss change trend are used as indicators. When the indicators stabilize at the expected level, the model is determined to be converged, and the GAT-PPO model for generating a scheduling action sequence for a hop scheduling granularity through a Critic-Actor double-flow decision network using a PPO algorithm is obtained; A scheduling action sequence for a hop scheduling granularity is generated through the GAT-PPO model based on the node features; The scheduling action sequence is encoded as a GA chromosome, the TT flow time slot for TT flow transmission is obtained through a genetic algorithm, and the initialized blank gating table is updated based on the TT flow time slot to obtain a target GCL table; According to the dynamic characteristics of the AVB flow, the priority of the AVB flow is determined through the APRA algorithm; According to the priority, the AVB flow is scheduled based on the non-TT flow time slot determined by the target GCL table through a credit value shaper using the MQSA algorithm, and multi-business flow scheduling is realized.

2. The hybrid intelligent-driven multi-traffic flow deterministic scheduling method of claim 1, wherein, The step of encoding the scheduling action sequence as a chromosome of GA, obtaining TT stream time slots of TT stream transmission by genetic algorithm, and updating the initialized blank gating table based on the TT stream time slots to obtain a target GCL table, comprises: encoding the scheduling action sequence as a chromosome matrix, wherein the row index corresponds to the switching node, the column index corresponds to the time slot, and the gene bit value is the hop ID or the idle identifier; verifying the chromosome matrix, and genetically evolving the verified chromosome matrix to update the global optimal solution counter; when no improvement is made for Thre1 consecutive generations, outputting a feasible hop probability distribution based on the chromosome state, inserting a high-yield hop into the idle time slot of the initialized blank gating table under the condition of meeting the constraint condition, guiding the optimization direction through a reward signal, resetting the global optimal solution counter after completion, and obtaining an elite chromosome; storing the elite chromosome decoded as a state-action trajectory into an experience replay pool, and fine-tuning the Actor network parameters by using the elite sample; after reaching the maximum number of iterations, decoding the global optimal chromosome to generate a final GCL scheduling table.

3. The hybrid intelligent-driven multi-traffic flow deterministic scheduling method of claim 1, wherein, The step of determining the priority of the AVB stream by the APRA algorithm according to the dynamic characteristics of the AVB stream, comprises: when the AVB stream arrives, calculating the initial expected priority of the AVB stream based on the deadline of the AVB stream, and setting a default priority adjustment parameter for the AVB stream; at the end of each time slot, calculating the current demand rate and the expected completion time of the AVB stream; calculating the utility value of the current time slot according to the current demand rate and the expected completion time; comparing the utility value changes of adjacent time slots, determining the utility gradient, and updating the priority adjustment parameter according to the utility gradient; based on the updated priority adjustment parameter, the expected priority of the AVB stream is probabilistically mapped to the actual priority.

4. The hybrid intelligent-driven multi-traffic flow deterministic scheduling method of claim 1, wherein, The step of scheduling the AVB stream in the non-TT stream time slot determined according to the target GCL table by the credit value shaper based on the MQSA algorithm to realize multi-service flow scheduling according to the priority, comprises: determining whether the current time slot is a non-TT stream transmission time slot according to the target GCL table, if the current time slot is a non-TT stream transmission time slot, determining whether CreditA under the current time slot is greater than zero: if CreditA under the current time slot is greater than zero, determining whether the current queue is empty, if it is empty, setting CreditA to zero; if the current queue is not empty, determining whether the AVB stream with the highest priority can be scheduled, if it can be scheduled, assigning the current time slot to the AVB stream with the highest priority, and updating CreditA and CreditB in the transmission process through the credit value shaper; if CreditA under the current time slot is less than or equal to zero, determining whether CreditB under the current time slot is greater than zero; if CreditB under the current time slot is greater than zero, determining whether the current queue is empty, if it is empty, setting CreditB to zero; If the current queue is not zero, it is determined whether the highest priority AVB stream can be scheduled, and if so, the highest priority AVB stream is assigned the current time slot, and CreditA and CreditB are updated during transmission by a credit shaper; If CreditB under the current time slot is less than or equal to zero, a time slot is assigned to the BE stream according to the FIFO principle.

5. A hybrid intelligent driven multi-traffic flow deterministic scheduling apparatus, characterized in that, The device comprises: An initialization module configured to calculate a superframe period by a centralized network configuration controller according to service flow requirements and a clock synchronization signal, and establish an initialized blank gating table based on the superframe period; The constructing module is configured to model the network topology information and the traffic state of the TT flow into a graph structure, extract node features of the graph structure using a graph attention network, and generate a scheduling action sequence for a hop scheduling granularity based on the node features by using a proximal policy optimization algorithm; the step of modeling the network topology information and the traffic state of the TT flow into a graph structure, extracting node features of the graph structure using a graph attention network, and generating a scheduling action sequence for a hop scheduling granularity based on the node features by using a proximal policy optimization algorithm comprises: modeling the network topology information and the traffic state of the TT flow into a time-varying graph structure, capturing the space-time dependence relationship between flow hop nodes using a graph attention network with a multi-head bidirectional attention mechanism, generating a five-dimensional state representation covering transmission delay, remaining hop count, scheduling label, port flow sending time and hop completion time as node features through feature aggregation and gating update; constructing a low-altitude TSN network environment including dynamic unmanned aerial vehicle nodes, wireless link models and typical service flow scenarios in a simulation platform, and creating a GAT encoder with a multi-phase double attention head mechanism, Actor policy network parameters and Critic value network; based on the current environment state, the GAT encoder generates high-order state representations according to the five-tuple features of each unmanned aerial vehicle node for the TT flow and the topological edge relationship; the old Actor network in the PPO algorithm generates actions by probability sampling, obtains immediate rewards and the next state after the actions interact with the environment, and then generates complete trajectories, and stores the complete trajectories in an experience replay buffer; trajectory data is collected from the buffer, the PPO agent loss is calculated through the high-order state representations, the GAT, Actor and Critic network parameters are iteratively updated using gradient descent, and the training is completed after a predetermined number of rounds, and the average round reward, round length and Critic loss are used as indicators; when the indicators stabilize at the expected level, it is determined that the model converges, and a GAT-PPO model for generating a scheduling action sequence for a hop scheduling granularity by using a Critic-Actor double-flow decision network and a PPO algorithm is obtained; a scheduling action sequence for a hop scheduling granularity is generated based on the node features by using the GAT-PPO model. The encoding module is configured to encode the scheduling action sequence as a chromosome of a GA, obtain TT stream time slots for TT stream transmission by using a genetic algorithm, and update the initialized blank gating table based on the TT stream time slots to obtain a target GCL table; The priority determination module is configured to determine the priority of the AVB stream according to a dynamic characteristic of the AVB stream by using an APRA algorithm. The scheduling module is configured to schedule the AVB stream in non-TT stream time slots determined according to the target GCL table by using a credit shaper based on an MQSA algorithm according to the priority, so as to implement multi-service stream scheduling.

6. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, The processor implements the hybrid intelligent driven multi-service stream deterministic scheduling method according to any one of claims 1 to 4 when executing the computer program.

7. A computer-readable storage medium storing a computer program, wherein the computer program comprises the following steps of: receiving a request for a resource from a client; determining whether the client is authorized to access the resource; and if the client is authorized to access the resource, providing the resource to the client. The computer program is executed by the processor to implement the hybrid intelligent driven multi-service stream deterministic scheduling method according to any one of claims 1 to 4.

Images