Round-trip delay sensing type ground-non-ground fusion link intelligent scheduling method for multi-path transmission

Abstract

The invention discloses a multi-path transmission-oriented round-trip delay sensing type ground-non-ground fusion link intelligent scheduling method, and belongs to the technical field of communication networks. The method comprises the following steps: firstly, sensing and collecting state parameters such as RTT, packet loss rate, bandwidth and the like of a plurality of ground and non-ground links in real time; secondly, performing feature extraction, normalization and time sequence modeling on the link data to form uniform link quality representation; then, constructing a scoring mechanism based on RTT perception, and outputting a scheduling priority score for each link through a heuristic function or a learning model; then, main path selection or multi-path weighted data distribution is executed according to a scoring result and a service requirement; and finally, monitoring a scheduling result through a feedback closed loop, and updating scoring model parameters online to adapt to network changes. According to the method, intelligent link scheduling is realized by deeply utilizing dynamic indexes such as RTT, the link utilization rate in a space-ground fusion network is remarkably improved, and the method is suitable for multi-path transmission in a high-speed mobile scene.

Description

Technical Field

[0001] This invention belongs to the field of communication network technology, and in particular relates to an intelligent scheduling method for terrestrial-non-terrestrial network (TN-NTN) converged links with round-trip time (RTT) awareness for multipath transmission. It belongs to the interdisciplinary research direction of satellite communication and terrestrial cellular network convergence, intelligent scheduling and control, and multipath transmission optimization. Background Technology

[0002] In the architecture of 6G mobile communication networks, non-terrestrial networks (NTN) are considered an important supplement and extension to terrestrial networks (TN). The integrated space-ground communication system (i.e., the TN-NTN converged network) built by the integration of the two is gradually becoming an important communication support for wide-area, airspace, and extreme environments in the future. This converged network achieves wide coverage, high resilience, and strong disaster recovery capabilities through low Earth orbit (LEO), medium Earth orbit (MEO), geostationary Earth orbit (GEO), and high altitude platform stations (HAPS), and is widely applicable to key scenarios such as emergency communication, marine communication, network coverage in remote areas, and intercontinental communication.

[0003] To further improve link utilization efficiency, end-to-end throughput performance, and service continuity, the Multipath Transmission Control Protocol (MPTCP) has gradually become an important technical means to support load balancing and link redundancy in terrestrial-space converged networks. However, existing multipath scheduling algorithms are mostly based on static parameters or historical experience of terrestrial networks for path selection, making it difficult to accurately adapt to the characteristics of NTN links, such as large dynamic latency fluctuations, bandwidth limitations, and transmission instability. In the complex environment of TN-NTN heterogeneous paths coexisting, if the scheduling mechanism responds slowly to changes in network status, it can easily lead to congestion on one link while another is idle, thereby reducing transmission efficiency, causing latency jitter, and even packet out-of-order delivery.

[0004] RTT, as a core indicator reflecting link transmission quality and latency overhead, is dynamically changing and significantly heterogeneous. LEO satellite links typically exhibit high RTT fluctuations, while terrestrial links tend to be more stable. Traditional scheduling mechanisms often only update RTT estimates during actual data transmission, lacking timely awareness of short-term fluctuations and link degradation. This causes scheduling algorithms to fail in rapidly changing link environments, making it impossible to achieve optimal resource allocation.

[0005] Furthermore, current RTT-aware scheduling schemes are limited to terrestrial MPTCP environments and have not yet explored their adaptation mechanisms in TN-NTN converged networks, incorporating factors such as track changes, congestion feedback, and link prediction. The lack of a systematic scheduling strategy for terrestrial-space converged scenarios, combining RTT estimation and multi-path joint optimization, has become a key bottleneck restricting the efficient operation of integrated terrestrial-space communication systems.

[0006] Therefore, there is an urgent need for an intelligent scheduling method that combines RTT dynamic estimation mechanism and multi-path scheduling logic to adapt to the characteristics of TN-NTN heterogeneous link environment, so as to achieve the coordinated utilization of space-ground link resources and service quality assurance. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention proposes an RTT-aware intelligent scheduling method for multi-path transmission in TN-NTN converged links. This method aims to solve problems such as strong link heterogeneity, scheduling lag, and unstable response in converged communication networks, thereby improving data transmission efficiency and link utilization in multi-path coexistence scenarios. This method is applicable to multi-path communication systems deployed in parallel with terrestrial and non-terrestrial networks, especially in converged environments including satellite and terrestrial links. The main steps include:

[0008] 1) Multi-link status awareness and data collection: When the terminal device is simultaneously connected to multiple TN and NTN links, it continuously collects communication characteristics such as RTT, bandwidth usage, packet loss rate, and queue length for each link. The collected information can come from protocol stack feedback, active probe packets, link layer statistics, etc.

[0009] 2) Link Quality Modeling and Feature Representation: Feature extraction, normalization, and time series modeling are performed on the collected raw link data to form a unified link quality vector representation. This representation is used to capture the instantaneous changes in link status and their temporal evolution trends, providing input for subsequent scheduling strategies.

[0010] 3) Construction of an RTT-aware scoring mechanism: Design a scoring function or model to output a scheduling priority score for each link based on the input link characteristics (focusing on RTT, congestion level, etc.). This model can be constructed using machine learning methods, such as lightweight neural networks, regression models, or heuristic functions. See the implementation example for details on the specific structure and training strategy.

[0011] 4) Link scheduling and task allocation strategy formulation: Based on the scoring results, combined with the service type (such as latency-sensitive, bandwidth-sensitive, etc.) and network policies (such as load balancing, primary / backup protection, etc.), the optimal path or data flow is selected or allocated. Multiple scheduling modes are supported, including primary path selection, alternative path preparation, and multi-path concurrent transmission.

[0012] 5) Scheduling result feedback and online model update: Real-time monitoring of scheduling execution results and network response status, and adjustment of scoring model parameters based on actual results to achieve online iterative optimization of the model to adapt to dynamic changes in link conditions and business fluctuations.

[0013] Furthermore, in the multi-link state awareness and data collection steps, the RTT is calculated using the data packet sending timestamp and the acknowledgment receiving time, and the packet loss rate is calculated using the total number of sent packets and the number of acknowledgment received packets.

[0014] Furthermore, the link quality modeling and feature representation steps include:

[0015] For each link, a feature set including RTT standard deviation, packet loss rate, available bandwidth, and queuing delay is constructed;

[0016] Perform max-min normalization on the features;

[0017] A time series feature matrix is ​​constructed based on a sliding window to preserve the dynamic evolution trend of the link state.

[0018] Furthermore, in the scoring mechanism construction step, the heuristic scoring function is:

[0019] Design a scoring function ,in The weights are adjustable, and i is 1, 2, or 3. For link normalized RTT, To normalize the packet loss rate of the link, To normalize the available bandwidth of the link, a heuristic scoring function is used to weight and fuse components such as RTT, packet loss, and congestion level to form a highly interpretable path score.

[0020] Furthermore, in the scoring mechanism construction step, the learning-based scoring model is a lightweight neural network, whose training objective is to minimize the difference between the predicted score and historical performance feedback, and the loss function is:

[0021] ;

[0022] in, For the true score of the k-th link, The coefficient of the regularization term, This serves as a constraint on model complexity.

[0023] Furthermore, the link scheduling and task allocation steps include:

[0024] Choose the path with the highest rating as the main path;

[0025] If the service supports multipath transmission, a weighted allocation strategy is used to allocate data streams;

[0026] Supports concurrent transmission or dynamic switching between the primary path and alternative paths;

[0027] It supports data distribution control at the packet level or the stream level.

[0028] Furthermore, the scheduling result feedback and online model update steps include:

[0029] Collect actual link performance data after scheduling;

[0030] Calculate the scoring bias and trigger model parameter updates;

[0031] If it is a learning model, then perform incremental training or fine-tuning;

[0032] If it is a heuristic model, the weights or thresholds are dynamically adjusted.

[0033] This method is applicable to terminal devices that integrate multiple network types (cellular, satellite, relay, high-altitude platforms, etc.), exhibiting excellent adaptability and robustness, particularly in high-speed mobile environments (such as vehicle-mounted, rail transit, and UAV platforms). By introducing an RTT-aware mechanism and a feature-based intelligent scheduling strategy, this invention achieves low-latency, high-reliability, and high-utilization data transmission in multi-link communication scenarios, effectively improving the service quality and system performance of converged networks.

[0034] The beneficial effects of this invention are as follows: Compared with existing technologies, the RTT-aware TN-NTN converged link intelligent scheduling method proposed in this invention for multi-path transmission has significant advantages. By introducing a multi-dimensional link state modeling mechanism based on RTT features, the scheduling engine's perception granularity and response speed to link quality changes in complex network environments are effectively improved. Compared with traditional scheduling methods based solely on static indicators (such as signal strength and static bandwidth) or fixed strategies, this invention, through real-time collection and dynamic modeling of link features, especially the deep utilization of latency indicators such as RTT, makes scheduling decisions more consistent with the actual link performance and service requirements. Furthermore, combined with the scoring model construction and online feedback mechanism, the scheduling strategy has adaptive capabilities, enabling rapid adjustment of the strategy when network conditions change, thereby avoiding performance degradation caused by link mutations. This method has good generalization ability and stability in scenarios such as multi-path concurrency, path selection, and primary / backup switching, and is particularly suitable for communication tasks with high real-time and reliability requirements in high-speed mobile terminal environments. Therefore, this invention has significant technical advantages in ensuring communication continuity, improving link utilization, reducing latency fluctuations, and enhancing the intelligence of multi-path scheduling, and has good prospects for engineering promotion and industrial application value. Attached Figure Description

[0035] Figure 1 Structure diagram of the multi-link state awareness and data collection module;

[0036] Figure 2 Flowchart of link quality modeling and feature representation;

[0037] Figure 3 Structure diagram of the RTT-aware scoring mechanism;

[0038] Figure 4 Logic diagram of link scheduling and task allocation strategy;

[0039] Figure 5 Flowchart of scheduling result feedback and online model update. Detailed Implementation

[0040] The following describes the RTT-aware TN-NTN converged link intelligent scheduling method proposed in this invention, oriented towards multi-path transmission, using a specific application scenario. This embodiment takes a converged terminal that simultaneously possesses terrestrial cellular communication links and low-Earth orbit satellite communication links as an example. This terminal operates in a protocol environment with multi-path transmission capabilities (such as MPTCP or MPQUIC), can maintain multiple heterogeneous communication links simultaneously, and achieves data distribution and path selection through a link scheduling mechanism. The specific steps are as follows:

[0041] Step 1: Multi-link Status Awareness and Data Collection Module

[0042] Multi-link status awareness and data collection steps: When the terminal device simultaneously accesses multiple terrestrial network (TN) links and non-terrestrial network (NTN) links, it continuously collects the round-trip time (RTT), packet loss rate, available bandwidth, queuing delay and channel quality indicators of each link.

[0043] like Figure 1 As shown, after establishing connections with multiple links, the terminal device first activates the link status awareness module to collect real-time status information of all available paths, including cellular links, LEO satellite links, and HAPS links. The main raw link parameters collected include:

[0044] Round-Trip Time (RTT) can be determined by the timestamps sent in the data packets. With ACK reception time The calculation shows that: ;

[0045] Packet Loss Rate (P) loss This can be estimated statistically by comparing the total number of packets sent with the number of acknowledgments received. ,in The total number of data packets sent. The number of confirmed packages received;

[0046] Available bandwidth B avail This is obtained using link-layer feedback or bandwidth probing algorithms;

[0047] Queue length or queue delay (D q ): Estimated through buffer state;

[0048] Channel quality metrics (such as RSSI, SINR, etc.).

[0049] The above parameters are obtained from the following sources: 1) Protocol feedback (TCP ACK delay, window sliding information); 2) Control layer or application layer probe packets (such as Ping or timed heartbeat packets); 3) MAC / PHY layer feedback (buffer status, rate changes, etc.).

[0050] All raw data is stored in a unified "link state database", forming a time-series structure as follows: ,in This represents the historical state sequence of the k-th link. For time points The observed feature vectors include the main original link parameters mentioned above.

[0051] Step 2: Link Quality Modeling and Feature Representation

[0052] Link quality modeling and feature representation steps: extract features, normalize and perform time series modeling on the collected link status data to form a unified link quality vector representation;

[0053] like Figure 2 As shown, in order to extract a unified representation that can be used for scheduling from the multi-dimensional link features, this module performs the following processing steps on the state sequence of each link:

[0054] 1) Feature extraction: Construct a feature set for each link k. ,in For The standard deviation of the RTT of link k within a certain time window with the endpoint as the endpoint. Packet Loss Ratio represents the percentage of packets lost in a given context. The proportion of data packets lost on link k within a certain time window with the endpoint as the endpoint, out of the total number of data packets sent, is used to measure link reliability. Available Bandwidth represents the actual data transmission rate that link k can provide under the current network conditions, reflecting the transmission capacity of the link. QueueingDelay represents the additional delay introduced by link k during forwarding due to queuing, and is used to characterize link delay performance.

[0055] 2) Data normalization: Normalize the values ​​of all dimensions to the [0,1] interval to eliminate the influence of units. To ensure the comparability of each dimension, max-min normalization is used. ,in For the j-th feature, and These are the normalized upper and lower bounds, respectively;

[0056] 3) Time series modeling: Short-term time series vectors are constructed based on a sliding window to preserve the dynamic evolution trend of link states, and a sliding window matrix is ​​constructed. Where w is the length of the sliding window and t is the current time point. Let f be the time series feature matrix of the nth link. t The link feature vector at time t is composed of metrics such as round-trip time (RTT), RTT standard deviation, packet loss rate, available bandwidth, and queuing delay.

[0057] Step 3: Construction of an RTT-aware scoring mechanism

[0058] The steps for building an RTT-aware scoring mechanism are as follows: Based on the link quality vector, output the scheduling priority score for each link through a heuristic scoring function or a learning-based scoring model.

[0059] like Figure 3 As shown, the core task of the scoring module is to output a scheduling priority score based on the input link feature vector, which drives link selection and data allocation. The scoring mechanism can be implemented in one of the following two ways:

[0060] 1) Heuristic scoring function: Designing a scoring function ,in ( ) represents an adjustable weight. For the normalized RTT of link k, Let k be the normalized packet loss rate. The normalized available bandwidth of link k is used. A heuristic scoring function is employed to weight and fuse components such as RTT, packet loss, and congestion level to form a highly interpretable path score.

[0061] 2) Learning-based scoring model: Constructing a lightweight neural network ,in The parameter is Neural network model, It is the first k The time-series feature matrix of each link. A scoring model is trained using historical data to obtain a better decision boundary without increasing computational overhead. The training objective of the neural network is to minimize the difference between the predicted score and historical performance feedback; therefore, the loss function is constructed as follows:

[0062] ;

[0063] in, For the true score of the k-th link, The coefficient of the regularization term, This serves as a constraint on model complexity.

[0064] Step 4: Link Scheduling and Task Allocation Strategies

[0065] Link scheduling and task allocation steps: Based on the scoring results, combined with the service type and network policy, select the main path or allocate the data flow;

[0066] like Figure 4 As shown, after obtaining the score results for each path, the scheduling engine combines the current service type (such as Voice over IP, video streaming, and File Transfer Protocol FTP based on Internet Protocol) and policy rules to execute the following scheduling logic:

[0067] 1) Select the path with the highest rating as the main path: Where k* is the index of the selected optimal path, corresponding to the candidate path with the highest score.

[0068] 2) If the business can tolerate multiple paths, adopt a weighted allocation strategy: ,in Assign weights to the data in each link;

[0069] 3) Supports concurrent allocation strategies for primary and alternative paths (e.g., 70% for primary paths and 30% for alternative paths), or dynamic switching based on path availability;

[0070] 4) Supports data distribution control at the packet level or stream level (configurable according to system capabilities);

[0071] 5) During the execution of the scheduling strategy, the distribution decision will be sent to the protocol stack and the scheduling results will be recorded in real time.

[0072] Step 5: Scheduling Result Feedback and Online Model Update

[0073] Scheduling result feedback and online model update steps: Monitor scheduling results in real time, adjust scoring mechanism parameters based on actual link performance, and achieve online model optimization.

[0074] like Figure 5 As shown, to improve the adaptability of scheduling strategies to network dynamics, this invention introduces a scheduling feedback mechanism and an online learning module. After scheduling is implemented, the terminal continues to track actual link performance and updates the scoring mechanism based on feedback. Specifically, this includes:

[0075] Collect the actual link performance after scheduling (such as actual RTT, packet loss rate, and packet out-of-order rate). ;

[0076] Calculate the scoring bias to form the model error;

[0077] If the scoring model is a learned architecture, the parameters are updated based on error-triggered fine-tuning or incremental training. : ,in For learning rate, The gradient of the loss function with respect to the model parameters θ represents the direction and magnitude of the parameter updates; L (loss function) is used to measure the difference between the model output and the true result.

[0078] If it is a heuristic model, the weight parameters or thresholds are dynamically adjusted based on the statistical results.

[0079] Through a feedback loop mechanism, the scoring model and scheduling engine can continuously adapt to changes in network state, improving long-term performance stability.

Claims

1. A round-trip delay-aware intelligent scheduling method for multi-path transmission ground-non-ground fusion links, characterized in that, Includes the following steps: Multi-link status awareness and data collection steps: When the terminal device simultaneously accesses multiple terrestrial network (TN) links and non-terrestrial network (NTN) links, it continuously collects the round-trip time (RTT), packet loss rate, available bandwidth, queuing delay, and channel quality indicators of each link. Link quality modeling and feature representation steps: extract features, normalize and perform time series modeling on the collected link status data to form a unified link quality vector representation; The steps for building an RTT-aware scoring mechanism are as follows: Based on the link quality vector, output the scheduling priority score for each link through a heuristic scoring function or a learning-based scoring model. Link scheduling and task allocation steps: Based on the scoring results, combined with the service type and network policy, select the main path or allocate the data flow; Scheduling result feedback and online model update steps: Monitor scheduling results in real time, adjust scoring mechanism parameters based on actual link performance, and achieve online model optimization.

2. The method according to claim 1, characterized in that, In the multi-link state awareness and data collection steps, the RTT is calculated using the data packet sending timestamp and the acknowledgment receiving time, and the packet loss rate is calculated by the total number of sent packets and the number of acknowledgment received packets.

3. The method according to claim 1, characterized in that, The link quality modeling and feature representation steps include: For each link, a feature set including RTT standard deviation, packet loss rate, available bandwidth, and queuing delay is constructed; Perform max-min normalization on the features; A time series feature matrix is ​​constructed based on a sliding window to preserve the dynamic evolution trend of the link state.

4. The method according to claim 1, characterized in that, In the scoring mechanism construction step, the heuristic scoring function is: Design a scoring function ,in The weights are adjustable, and i is 1, 2, or 3. For link normalized RTT, To normalize the packet loss rate of the link, To normalize the available bandwidth of the link, a heuristic scoring function is used to weight and fuse components such as RTT, packet loss, and congestion level to form a highly interpretable path score.

5. The method according to claim 4, characterized in that, In the scoring mechanism construction step, the learning-based scoring model is a lightweight neural network, whose training objective is to minimize the difference between the predicted score and historical performance feedback, and the loss function is: ; in, For the true score of the k-th link, The coefficient of the regularization term, This serves as a constraint on model complexity.

6. The method according to claim 1, characterized in that, The link scheduling and task allocation steps include: Choose the path with the highest rating as the main path; If the service supports multipath transmission, a weighted allocation strategy is used to allocate data streams; Supports concurrent transmission or dynamic switching between the primary path and alternative paths; It supports data distribution control at the packet level or the stream level.

7. The method according to claim 1, characterized in that, The scheduling result feedback and online model update steps include: Collect actual link performance data after scheduling; Calculate the scoring bias and trigger model parameter updates; If it is a learning model, then perform incremental training or fine-tuning; If it is a heuristic model, the weights or thresholds are dynamically adjusted.

Images