Non-ground network-oriented intelligent forward erasure coding (FEC) service flow scheduling method

Abstract

The invention discloses an intelligent forward erasure coding (FEC) service flow scheduling method for a non-ground network. The method comprises the following steps: constructing a transmission system model by adopting an FEC transmission architecture based on a performance enhancement agent; in a transmission system model, based on an intelligent scheduling mechanism of deep reinforcement learning, a scheduling problem is described in a formalized manner by constructing a Markov decision process, and a network state space, an action space and a reward function are defined; designing a Dueling-DQN decision model, combining the output of the state value function and the output of the action dominant function, and training based on a priority experience playback mechanism; and designing a self-adaptive scheduling strategy to realize self-adaptive resource allocation scheduling of different service flows in a high-load scene. According to the method, the network state can be sensed in real time, the scheduling strategy is dynamically adjusted, and network resources are distributed in an optimal mode. The end-to-end ordered delivery time delay can be obviously reduced, the throughput stability is improved, and the overall optimization of the end-to-end transmission performance is realized.

Description

Technical Field

[0001] This invention belongs to the field of communication network technology, and particularly relates to an intelligent forward erasure coding (FEC) service flow scheduling method for non-terrestrial networks. Background Technology

[0002] Low-latency packet transmission technology is a crucial requirement in the current information and communication field, playing a vital role in applications such as telemedicine and ultra-high-definition live streaming. These services demand not only high throughput but also low latency and high reliability in data transmission. Loss or timeout of any data packet can lead to service interruption, severely impacting user experience and system reliability. However, the traditional TCP / IP network transmission model and retransmission-dependent error control mechanisms struggle to meet the stringent deterministic Quality of Service (QoS) requirements in long-latency, packet-loss-prone non-terrestrial network (NTN) environments.

[0003] To improve transmission reliability and reduce delivery latency, Forward Erasure Correction (FEC) is a viable technical approach. Its core principle is to calculate and add redundant packets to the original data block at the sending end. Once the receiving end successfully receives a sufficient number of data packets, it can recover the original information through decoding, effectively reducing retransmission requirements. However, in multi-service stream transmission environments using FEC, the system faces complex scheduling challenges: different service streams compete for limited network resources, and each stream has different requirements for latency, throughput, and reliability. Simultaneously, the system must balance sending original data packets with redundancy repair packets to ensure transmission reliability while avoiding unnecessary bandwidth overhead. These scheduling decisions collectively determine the effectiveness of FEC technology and directly affect the overall transmission performance of the system.

[0004] Traditional traffic flow scheduling strategies, such as polling-based scheduling algorithms (RR, DRR, etc.) and GPS-based scheduling algorithms (WFQ, WF2Q, etc.), are mostly based on predefined heuristic rules or static priority allocation. These methods perform well in constant service rates and stable network environments, providing relatively fair bandwidth allocation that conforms to priority rules. However, in dynamic network environments, especially when forward erasure coding is applied, such as when deployed as a performance enhancement proxy (PEP), the limitations of these traditional scheduling algorithms become significant. FEC transmission alters the characteristics of the transmitted traffic flow due to the introduction of redundant packet transmission during the encoding process; furthermore, the latency of the decoding process causes dynamic fluctuations in the equivalent service rate. Fixed-weight scheduling decisions struggle to achieve adaptive optimal resource allocation among multiple traffic flows. Summary of the Invention

[0005] Purpose of the invention: The purpose of this invention is to provide an intelligent forward erasure coding (FEC) service flow scheduling method for non-terrestrial networks (NTN) and complex terrestrial networks, in order to solve the problems that traditional scheduling algorithms (such as DRR, WFQ, etc.) are difficult to adapt to dynamic traffic and cannot simultaneously guarantee low latency, high reliability and fairness of multiple service flows in long-latency and packet-loss-prone environments.

[0006] Technical solution: The present invention provides an intelligent forward erasure coding (FEC) service flow scheduling method for non-terrestrial networks, comprising the following steps:

[0007] Step 1: Use the FEC transmission architecture based on performance-enhanced proxies to build the transmission system model;

[0008] Step 2: In the transmission system model, an intelligent scheduling mechanism based on deep reinforcement learning is used to formally describe the scheduling problem by constructing a Markov decision process, and the network state space, action space and reward function are defined.

[0009] Step 3: Design the Dueling-DQN decision model, combine the outputs of the state value function and the action advantage function, and train it based on the priority experience replay mechanism;

[0010] Step 4: Design an adaptive scheduling strategy to achieve adaptive resource allocation and scheduling for different business flows under high load scenarios.

[0011] Furthermore, step 1 specifically involves the following: The transmission system model adopts an FEC transmission architecture based on Performance Enhancement Proxy (PEP). The sending end buffers data packets from multiple service flows into corresponding input queues, with each queue corresponding to an independent service flow. The scheduler selects data packets from each queue at fixed intervals as source packets and sends them to the FEC encoding module. The encoding module, based on stream coding technology, inserts redundant repair packets into the source packet sequence at a preset code rate. The receiving end decodes the packets using instantaneous Gaussian elimination to recover lost data packets, thereby avoiding the high latency caused by traditional retransmission mechanisms. The system monitors scheduling latency and end-to-end delivery latency in real time to dynamically optimize the scheduling strategy.

[0012] Furthermore, step 2 specifically includes the following steps: Step 2.1: Define the network state space. The system state vector contains flow-level and global-level information. For each service flow at the current scheduling time t, the state vector contains the length of the service flow's waiting queue. Head-of-line data packet waiting time Decoding window width (DWW), service flow weights and the length of undelivered business flows in PEP Finally, its status was determined. ; Step 2.2: Define the action space. The system adopts a D+1 discrete action space design, where the first D actions correspond to allocating the current scheduling opportunity to the corresponding service flow, and the D+1th action is to skip scheduling and is used to control the queue length of the FEC encoder; given the state The Dueling-DQN module selects an action. The action selection strategy integrates multiple optimizations, taking into account the real-time status of each business flow and the requirements of the FEC mechanism. During the encoding and decoding process, an adaptive skip mechanism based on window width is adopted. When the encoding window exceeds the threshold, the probability of selecting to skip the scheduling, i.e., the D+1th action, is adaptively increased to prevent queue overflow. Step 2.3: Define the reward function and design a comprehensive reward function based on multi-dimensional performance indicators: ; Based on the waiting latency of the header data packets of each service flow Current queue length and its priority weight The first step in the construction process is to design a latency ratio bonus to address the latency ratio control requirements between business flows. : ; ; In the formula, there are a total of D business flows. , Let i be the waiting time for the header data packet of the i-th traffic flow. Let G be the weight of the i-th business flow, and its range is... G=0 represents perfect fairness, and G=1 represents perfect unfairness. This is determined by comparing the current... Compared with the previous time step The value is used to evaluate the effectiveness of the scheduling actions. First, it is a constant; second, to maintain queue stability and prevent queue overflow, a queue overflow prevention reward is designed. The formula is as follows: ; In the formula This is the queue buffer threshold. Let be the length of the waiting queue for the i-th business flow. It is a constant; finally, to meet the priority requirements of each business flow, the priority is designed to be reward-oriented. First, calculate its weighted reference latency. : ; Then, a priority-biased reward is defined, with the following formula: ; In the formula , , , It is a constant. .

[0013] Furthermore, in step 3, the outputs of the state space and action space advantage functions are combined. Specifically, the system adopts a Dueling-DQN architecture, starting with one or more convolutional or fully connected layers to extract features of the input state. The parameters of the convolutional or fully connected layers are shared, and after the shared layer, the system splits into two paths, which are used to calculate the state value function and the action advantage function, respectively. The first path is used to calculate the state value function. The first path outputs a value representing the expected reward in a given state; the second path is used to calculate the action advantage function. It outputs a vector representing the advantage of taking each action in a given state; using a combination operation, it combines the output of the state-value function. and the output of the action advantage function Combining the two, the Q-network model Defined as: ; In the formula For a set of actions, The number of selectable actions; A summation index variable representing the current state of the action. Lower Action Space Every possible action in it; It is a state-value function, representing the expected cumulative reward for following the current policy in state s; This represents the expected reward gained by choosing action a in state s compared to choosing the average action. Represented as: .

[0014] Furthermore, in step 3, the training based on the priority experience replay mechanism specifically involves: the agent employing the priority experience replay mechanism to sample experience according to the TD error priority, thereby improving learning efficiency; the agent adopting a dynamically adjusted hyperparameter strategy, exploring the rate... The decay occurs during training, and its changing patterns can be described as follows: ; In the formula To train step count, Switch the threshold for the attenuation strategy; This is the minimum decay value in the early stage. The initial exploration rate, This is the initial decay factor; This represents the minimum decay value in the later stages. This is the minimum decay factor in the later stage; simultaneously, the importance sampling weights... Parameters vary with training steps Increasing to balance the bias of prioritizing experience replay, the changing patterns of which are described as follows: ; In the formula The growth coefficient, Importance sampling weights The initial value.

[0015] Furthermore, step 4 specifically involves: dynamically adjusting the scheduling strategy by real-time monitoring of link packet loss rate, queue status, and service weight to achieve adaptive resource allocation for different business flows under high load scenarios; continuously updating the Q-value table during the learning process, and combining the ε-greedy strategy to balance exploration and utilization, so that the system maintains stable end-to-end performance under sudden traffic and random packet loss conditions.

[0016] The present invention also discloses a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method of the present invention.

[0017] The present invention also discloses a computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implements the steps of the method of the present invention.

[0018] The present invention also discloses a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the method of the present invention.

[0019] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:

[0020] 1. This invention, based on a forward erasure coding transmission system model, utilizes a deep reinforcement learning algorithm to design and construct an intelligent adaptive service flow scheduling mechanism. This mechanism can perceive network status in real time and dynamically adjust scheduling strategies to allocate network resources optimally. Experiments show that this invention can significantly reduce end-to-end ordered delivery latency and improve throughput stability in high-load, lossy NTN and similar links. While ensuring the quality of real-time application data transmission, it effectively maintains the fairness of resource allocation among different service flows, achieving overall optimization of end-to-end transmission performance.

[0021] 2. Significantly reduce end-to-end latency: In non-terrestrial network environments with high latency and packet loss, the scheduling mechanism based on deep reinforcement learning proposed in this invention can dynamically perceive the network status and intelligently select the optimal scheduling strategy. Compared with traditional scheduling algorithms (such as DRR and WFQ), it can significantly reduce end-to-end ordered delivery latency under the same load conditions, and its advantages are more obvious in high load scenarios.

[0022] 3. Improve throughput stability and link utilization: By combining forward erasure coding and intelligent scheduling mechanisms, the system can maintain high throughput stability under conditions of sudden traffic and random packet loss, effectively improving link bandwidth utilization and avoiding performance fluctuations caused by frequent retransmissions.

[0023] 4. Ensure fairness and service quality among multiple service flows: The designed multi-objective reward function comprehensively considers latency, queue stability and priority weight, ensuring that high-priority service flows obtain lower latency while maintaining fairness among service flows and meeting differentiated service needs.

[0024] 5. Excellent environmental adaptability: The deep reinforcement learning-based scheduler has online learning and dynamic optimization capabilities, enabling it to adapt to dynamic environments such as sudden network traffic surges and parameter changes. It exhibits strong robustness and application potential, making it suitable for future high-dynamic non-terrestrial networks and complex terrestrial network scenarios. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of a service flow QoS-aware scheduling framework based on DRL under elastic erasure coding.

[0026] Figure 2 This is a flowchart of the workflow of a business flow processing intelligent agent based on deep reinforcement learning.

[0027] Figure 3 This is a diagram of the ns-3 and deep reinforcement learning integration framework based on shared memory.

[0028] Figure 4 The scheduling delay diagrams for DRR, WFQ, and DRL algorithms are shown under a priority ratio of 1:1.

[0029] Figure 5 Delivery delay for DRR, WFQ and DRL algorithms under priority 1:1.

[0030] Figure 6 The scheduling delay diagrams for DRR, WFQ, and DRL algorithms are shown for a priority ratio of 1:2.

[0031] Figure 7 The delivery delay diagrams for DRR, WFQ, and DRL algorithms are shown under a priority ratio of 1:2. Detailed Implementation

[0032] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings and specific embodiments. It should be noted that the described examples and technical solutions are only some embodiments of the present invention, and not all embodiments. The following embodiments or drawings are used to illustrate the present invention and are not intended to limit the scope of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0033] This invention proposes an intelligent FEC service flow scheduling algorithm for non-terrestrial networks, which specifically includes the following steps:

[0034] This invention considers the following: Figure 1 The diagram illustrates the architecture of a service flow scheduling system. This architecture assumes there are D service flows in the system, each corresponding to an arrival buffer queue, denoted as [Duplicate]. Data packets in each queue are numbered according to their arrival order, and the packet length is fixed at K bytes. Packets are processed at a rate... Enter the corresponding queue , Waiting for scheduling. To facilitate convenient and scalable deployment of FEC, performance-enhanced proxies are generally used in practical systems. The deployment method does not affect the specific design and analysis of the scheduling algorithm; therefore, this paper assumes that FEC is deployed using PEP. The scheduler, based on a specific scheduling strategy, ... Scheduling interval (i.e., rate) The data packet at the head of each queue is selected as the source packet and enters the FEC encoder queue deployed in PEP form. This invention uses FEC encoding for transmission between the sending and receiving ends, that is, repair packets are inserted at equal intervals in the source packet sequence. The receiving end then performs forward erasure decoding using instantaneous Gaussian elimination based on the received source packets and repair packets, thereby recovering packets that may have been lost during transmission.

[0035] Based on this, such as Figure 1 As shown, this invention examines two key latency indicators: one is scheduling latency (… ), that is, grouping by Arrive at the corresponding queue The waiting time experienced by a device being scheduled into the encoder reflects the timeliness and fairness of the scheduler's processing of each service flow; secondly, the end-to-end delivery delay. This encompasses the entire process of scheduling, sending, transmission, and receiving confirmation, directly determining the final latency performance of the service experience. During this period, This represents the time when the PEP sender completes encoding and sends the packet to the receiver. This indicates the moment when the PEP receiver receives the data packet.

[0036] The arrival model of this invention follows a Poisson-Pareto Burst Process (PPBP). The arrival process of each traffic flow is characterized by the burst interval (following an exponential distribution) and the burst duration (following a Pareto distribution), and its self-similarity is adjusted by the Hurst parameter. Furthermore, dynamic parameters are considered, that is, during transmission, the burst arrival rate Hurst parameter can vary randomly within a specific range to reflect the long-term correlation, burst clustering, and time-varying characteristics of network traffic.

[0037] Based on the above transmission model, the D service flows are assigned different priority weights according to their quality of service requirements. The system utilizes a deep reinforcement learning model to perceive the state of the network and each service flow in real time, and schedules groups from the waiting queues of each service flow to enter the FEC encoder when the scheduling time arrives.

[0038] The DRL scheduler of this invention is based on PEP and guides scheduling decisions by selecting the most suitable service flow scheduler for each time slot. For example... Figure 2 As shown, the model is based on the Dueling-DQN algorithm, which formalizes the scheduling decision problem into a Markov decision process to achieve adaptive learning and decision-making in complex network environments. The design of the states, actions, and reward functions used to train the Dueling-DQN model will be detailed below.

[0039] 1. State. The system's state representation includes information at both the flow level and the global level. For each service flow at the current scheduling time t, the state vector contains the length of the service flow's waiting queue. Head-of-line data packet waiting time Decoding window width (DWW), service flow weights and the length of undelivered business flows in PEP These key features ultimately determine its state. .

[0040] 2. Action Space. The system employs a D+1 discrete action space design. The first D actions correspond to allocating the current scheduling opportunity to the corresponding service flow, while the D+1th action is a skip scheduling action used to control the queue length of the FEC encoder. Given a state The Dueling-DQN module selects an action. This action selection strategy incorporates multiple optimizations, comprehensively considering the real-time status of each service flow and the requirements of the FEC mechanism. During the encoding and decoding process, an adaptive skip mechanism based on window width is employed. When a large encoding window is detected, the probability of selecting to skip the scheduled action (i.e., the D+1th action) is adaptively increased to prevent queue overflow.

[0041] 3. Reward Value. To accurately evaluate the scheduling actions selected by the agent, this invention designs the following comprehensive reward function based on multi-dimensional performance indicators:

[0042]

[0043] This function is primarily based on the waiting delay of the data packets at the head of each service flow queue. Current queue length Its priority weight The first step in the construction process is to design a latency ratio bonus to address the latency ratio control requirements between business flows. :

[0044]

[0045]

[0046] The range of G in the formula is [ 0 , 1 ] G=0 represents complete fairness, and G=1 represents complete unfairness. This is determined by comparing the current... Compared with the previous time step The value is used to evaluate the effectiveness of the scheduling actions. It is a constant. Secondly, to maintain queue stability and prevent queue overflow, a queue overflow prevention reward was designed. The formula is as follows:

[0047]

[0048] In the formula This is the queue buffer threshold. This is a constant. Finally, to meet the priority requirements of each business flow, a priority-biased reward system was designed. First, calculate its weighted reference latency. :

[0049]

[0050] Then, a priority-biased reward is defined, with the following formula:

[0051]

[0052] In the formula , , , It is a constant. .

[0053] This reward function integrates performance parameters from different dimensions to ensure that the agent can prioritize high-priority service flows when making scheduling decisions, while maintaining a reasonable latency ratio and preventing queue overflow, thereby achieving efficient utilization of network resources and quality of service assurance.

[0054] 4. Dueling-DQN Decision Model Design: The system adopts a Dueling-DQN architecture, starting with one or more convolutional or fully connected layers to extract features from the input state. The parameters of these layers are shared to effectively learn the representation of the input state. After the shared layers, the system splits into two paths for calculating the state-value function and the action advantage function, respectively. The first path is used to calculate the state-value function. The first path outputs a value representing the expected reward in a given state. The second path is used to calculate the action advantage function. It outputs a vector representing the advantage of taking each action given a state. It employs a combination operation, combining the output of the state-value function... and action advantage function The outputs are combined to form the Q-network model. Defined as:

[0055]

[0056] In the formula For a set of actions, This represents the number of selectable actions.

[0057] The agent employs a priority experience replay mechanism, sampling experience based on TD error priority to improve learning efficiency; the agent uses a dynamically adjusted hyperparameter strategy to increase the exploration rate. The decay occurs during training, and its changing patterns can be described as follows: ; In the formula To train step count, Switch the threshold for the attenuation strategy; This is the minimum decay value in the early stage. The initial exploration rate, This is the initial decay factor; This represents the minimum decay value in the later stages. This is the minimum decay factor in the later stage; simultaneously, the importance sampling weights... Parameters vary with training steps Increasing to balance the bias of prioritizing experience replay, the changing patterns of which are described as follows: ; In the formula The growth coefficient, Importance sampling weights The initial value.

[0058] This invention, relying on the ns-3 simulation environment and the PyTorch deep learning framework, designs and implements an integrated framework based on shared memory and a vectorized message interface. The core of this framework lies in using an efficient, low-latency bidirectional communication mechanism to transmit the real-time network state generated in the ns-3 simulation instance to the DRL environment on the Python side, and to receive the scheduling actions returned from it, thereby achieving closed-loop real-time control of the network scheduling strategy.

[0059] This framework is built on top of the ns3-ai module, such as Figure 3 As shown, this module provides the foundation for the interaction between ns-3 and the DRL model. This invention encapsulates communication content through well-defined data structures (Env and Act), ensuring the structured and type-safe exchange of data between the two parties. The framework's communication foundation is built upon a bidirectional vectorized message interface. This interface manages two shared memory vectors: Cpp2PyVector is responsible for transmitting state observation information from the C++ environment (ns-3) to the Python agent, while Py2CppVector is used to transmit action instructions generated by the agent in reverse. On the C++ side, the network sending application (ScpepSendApp) is the core agent of the simulation environment. It inherits from the application base class in ns-3 and embeds the interaction logic with the AI ​​model by overriding key scheduler functions. In this function, it collects real-time network state metrics (including queue lengths of each data stream, head packet latency, weights, and global encoding window state, etc.) and organizes this information into a normalized state vector. This state vector is populated into a predefined structure (Env), and then the structure data is written to a Cpp2PyVector in shared memory by calling the CppSendBegin and CppSendEnd methods of the message interface instance. After the write is complete, the interface triggers a synchronization signal to notify the Python side that new data has arrived.

[0060] On the Python side, a bound module (generated via Pybind11) provides direct access to shared memory. The DRL model reads the latest Env structure from Cpp2PyVector by calling the PyRecvBegin and PyRecvEnd methods of the interface, and extracts the state vector from it to input into its neural network. The agent calculates action decisions (e.g., choosing to schedule a certain business flow or not scheduling it) based on internal policies (such as -greedy exploration and exploitation), and writes the results into the corresponding fields of the Act structure. Finally, by calling the PySendBegin and PySendEnd methods, the Act structure containing the action instructions is submitted to Py2CppVector.

[0061] Upon detecting a new action, the C++-based ScpepSendApp immediately reads the Act structure from Py2CppVector, parses the action value, and performs the corresponding network operation (such as dequeuing data packets from the queue of a specified stream and sending them). The effect of this scheduling behavior will be reflected in the network state of the next simulation step, thus forming a complete "state-action-reward" loop that drives the continuous deep reinforcement learning process.

[0062] The entire communication process is built on shared memory, avoiding the overhead of serialization and network transmission, thus ensuring extremely low latency and high throughput in the interaction. Furthermore, the framework's structure is highly modular; state representation, reward functions, and agent algorithms are all implemented in Python. This allows researchers to flexibly experiment with different learning algorithms and problem modeling methods without recompiling the C++ simulation program, greatly improving the iterative efficiency and scalability of the experiments.

[0063] This invention has been extensively tested in ns-3 and compared with traditional scheduling algorithms, with the results as follows: Figures 4-7 As shown in the figure. Experimental results demonstrate that the scheduling method proposed in this invention significantly outperforms traditional algorithms in key metrics such as latency and fairness. Under different load scenarios, its average end-to-end latency and queue fluctuations are significantly reduced. Particularly in high-concurrency streaming scenarios, this method exhibits stronger stability and adaptability, validating the effectiveness and feasibility of deep reinforcement learning in dynamic resource scheduling. Figures 4-5 As shown, in scenarios with the same priority (1:1), the scheduling latency of the three algorithms is similar in the low-load range; when the load exceeds 300 packets / s, the latency of DRR and WFQ increases significantly, while DRL maintains a relatively stable growth trend. Regarding delivery latency, DRL is generally lower than traditional algorithms under medium to high loads, and its growth is more gradual, demonstrating better load adaptability and stability. Figures 6-7As shown, in scenarios with different priorities (1:2), DRR's service differentiation capability is weak, and its scheduling and delivery latency are both high. Both WFQ and DRR can effectively differentiate between high and low priority services, ensuring the service quality of high-priority flows. However, DRR can still maintain lower overall latency under high load, and the curve is more stable, indicating that it has better dynamic scheduling capabilities while maintaining service differentiation.

Claims

1. A smart forward erasure coding (FEC) service flow scheduling method for non-terrestrial networks, characterized in that, Includes the following steps: Step 1: Use the FEC transmission architecture based on performance-enhanced proxies to build the transmission system model; Step 2: In the transmission system model, an intelligent scheduling mechanism based on deep reinforcement learning is used to formally describe the scheduling problem by constructing a Markov decision process, and the network state space, action space and reward function are defined. Step 3: Design the Dueling-DQN decision model, combine the outputs of the state value function and the action advantage function, and train it based on the priority experience replay mechanism; Step 4: Design an adaptive scheduling strategy to achieve adaptive resource allocation and scheduling for different business flows under high load scenarios.

2. The intelligent forward erasure coding (FEC) service flow scheduling method for non-terrestrial networks according to claim 1, characterized in that, Step 1 is as follows: The transmission system model adopts an FEC transmission architecture based on performance-enhanced proxy (PEP). The sending end buffers data packets from multiple service flows into corresponding input queues. Each queue corresponds to an independent service flow. The scheduler selects data packets from each queue at fixed intervals as source packets and sends them to the FEC encoding module. The encoding module is based on stream coding technology and inserts redundant repair packets into the source packet sequence at a preset code rate. The receiving end decodes the data packets using instantaneous Gaussian elimination to recover lost data packets, thereby avoiding the high latency caused by traditional retransmission mechanisms. The system monitors the scheduling latency and end-to-end delivery latency in real time to dynamically optimize the scheduling strategy.

3. The intelligent forward erasure coding (FEC) service flow scheduling method for non-terrestrial networks according to claim 1, characterized in that, Step 2 specifically includes the following steps: Step 2.1: Define the network state space. The system state vector contains flow-level and global-level information. For each service flow at the current scheduling time t, the state vector contains the length of the service flow's waiting queue. Head-of-line data packet waiting time Decoding window width (DWW), service flow weights and the length of undelivered business flows in PEP Finally, its status was determined. ; Step 2.2: Define the action space. The system adopts a D+1 discrete action space design, where the first D actions correspond to allocating the current scheduling opportunity to the corresponding service flow, and the D+1th action is to skip scheduling and is used to control the queue length of the FEC encoder; given the state The Dueling-DQN module selects an action. The action selection strategy integrates multiple optimizations, taking into account the real-time status of each business flow and the requirements of the FEC mechanism. During the encoding and decoding process, an adaptive skip mechanism based on window width is adopted. When the encoding window exceeds the threshold, the probability of selecting to skip the scheduling, i.e., the D+1th action, is adaptively increased to prevent queue overflow. Step 2.3: Define the reward function and design a comprehensive reward function based on multi-dimensional performance indicators: ； Based on the waiting latency of the header data packets of each service flow Current queue length and its priority weight The first step in the construction process is to design a latency ratio bonus to address the latency ratio control requirements between business flows. : ； ； In the formula, there are a total of D business flows. , Let i be the waiting time for the header data packet of the i-th traffic flow. Let G be the weight of the i-th business flow, and its range is... G=0 represents perfect fairness, and G=1 represents perfect unfairness. This is determined by comparing the current... Compared with the previous time step The value is used to evaluate the effectiveness of the scheduling actions. First, it is a constant; second, to maintain queue stability and prevent queue overflow, a queue overflow prevention reward is designed. The formula is as follows: ； In the formula This is the queue buffer threshold. Let be the length of the waiting queue for the i-th business flow. It is a constant; finally, to meet the priority requirements of each business flow, the priority is designed to be reward-oriented. First, calculate its weighted reference latency. : ； Then, a priority-biased reward is defined, with the following formula: ； In the formula , , , It is a constant. .

4. The intelligent forward erasure coding (FEC) service flow scheduling method for non-terrestrial networks according to claim 1, characterized in that, In step 3, the outputs of the state space and action space advantage functions are combined. Specifically, the system adopts a Dueling-DQN architecture, starting with one or more convolutional or fully connected layers to extract features of the input state. The parameters of the convolutional or fully connected layers are shared. After the shared layer, the system splits into two paths, one for calculating the state value function and the other for calculating the action advantage function. The first path is used to calculate the state value function. The first path outputs a value representing the expected reward in a given state; the second path is used to calculate the action advantage function. It outputs a vector representing the advantage of taking each action in a given state; using a combination operation, it combines the output of the state-value function. and the output of the action advantage function Combining the two, the Q-network model Defined as: ； In the formula For a set of actions, The number of selectable actions; A summation index variable representing the current state of the action. Lower Action Space Every possible action in it; It is a state-value function, representing the expected cumulative reward for following the current policy in state s; This represents the expected reward gained by choosing action a in state s compared to choosing the average action. Represented as: 。 5. The intelligent forward erasure coding (FEC) service flow scheduling method for non-terrestrial networks according to claim 1, characterized in that, In step 3, the training based on the priority experience replay mechanism specifically involves: the agent adopting the priority experience replay mechanism to sample experience according to the TD error priority, thereby improving learning efficiency; the agent employing a dynamically adjusted hyperparameter strategy, with an exploration rate... The decay occurs during training, and its changing patterns can be described as follows: ； In the formula To train step count, Switch the threshold for the attenuation strategy; This is the minimum decay value in the early stage. The initial exploration rate, This is the initial decay factor; This represents the minimum decay value in the later stages. This is the minimum decay factor in the later stage; simultaneously, the importance sampling weights... Parameters vary with training steps Increasing to balance the bias of prioritizing experience replay, the changing patterns of which are described as follows: ； In the formula The growth coefficient, Importance sampling weights The initial value.

6. The intelligent forward erasure coding (FEC) service flow scheduling method for non-terrestrial networks according to claim 1, characterized in that, Step 4 specifically involves: dynamically adjusting the scheduling strategy by real-time monitoring of link packet loss rate, queue status, and service weight to achieve adaptive resource allocation for different business flows under high load scenarios; continuously updating the Q-value table during the learning process, and combining the ε-greedy strategy to balance exploration and utilization, so that the system maintains stable end-to-end performance under sudden traffic and random packet loss conditions.

7. A computer device comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method of claim 1.

8. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method of claim 1.

9. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method of claim 1.

Images