Low-latency data transmission scheduling method for 5G edge computing

Abstract

This invention discloses a low-latency data transmission scheduling method for 5G edge computing, relating to the field of 5G edge computing technology. The method includes acquiring a set of tasks reported by edge computing nodes, including the source node identifier, target node identifier, data size, and task generation timestamp for each transmission task; classifying tasks into latency-critical and latency-tolerant subsets based on preset latency thresholds and timestamps; reserving channels for latency-critical tasks based on predicted network state derived from historical channel quality and real-time load, and performing dynamic aggregation batch transmission for latency-tolerant tasks based on channel idle windows and data block sizes; and adjusting the channel reservation strategy during transmission according to actual latency deviations. This method can match different latency requirements, reduce latency fluctuations, improve channel utilization, and reduce transmission overhead.

Description

Technical Field

[0001] This invention belongs to the field of 5G edge computing technology, specifically a low-latency data transmission scheduling method for 5G edge computing. Background Technology

[0002] In 5G edge computing scenarios, data transmission scheduling is an important technology to ensure the efficiency of data interaction between edge nodes. Existing technologies mostly adopt a unified scheduling method, which centrally allocates all data transmission tasks reported by multiple edge computing nodes. It simply schedules based on the data volume or the current channel load, without distinguishing the latency requirements of different tasks.

[0003] In existing technical solutions, channel allocation mostly adopts static reservation or random allocation modes. The static reservation mode cannot be dynamically adjusted according to the network status, which easily leads to the waste of channel resources. The random allocation mode is difficult to guarantee the transmission performance of tasks with high latency requirements. At the same time, existing technologies do not combine historical channel quality and real-time network load to predict network status, nor do they perform differentiated scheduling for tasks with different latency requirements. As a result, latency-critical tasks are prone to exceeding latency limits, and latency-tolerant tasks cannot make full use of idle channel resources. The overall rationality and efficiency of transmission scheduling are insufficient.

[0004] Existing technologies cannot achieve accurate classification and scheduling based on task latency requirements, nor can they optimize channel allocation through network state prediction. They also cannot balance the latency requirements of latency-critical tasks with the resource utilization efficiency of latency-tolerant tasks. Therefore, a transmission scheduling method that can specifically solve the above problems is needed. Summary of the Invention

[0005] This invention aims to solve at least one of the technical problems existing in the prior art;

[0006] To this end, the present invention proposes a low-latency data transmission scheduling method for 5G edge computing, comprising:

[0007] Obtain the set of data transmission tasks reported by multiple edge computing nodes within the current transmission cycle. The set of data transmission tasks includes the source node identifier, target node identifier, data size, and task generation timestamp for each transmission task.

[0008] The data transmission task set is classified by latency sensitivity to generate a latency-critical task subset and a latency-tolerant task subset. The latency sensitivity classification is calculated based on a preset latency threshold and the task generation timestamp.

[0009] For the latency-critical subset of tasks, channel reservation processing based on predicted network state is performed, wherein the predicted network state is derived from historical channel quality indicators and real-time network load data;

[0010] For the latency-tolerant subset of tasks, a batch transmission scheduling process based on dynamic aggregation of available channels is executed, wherein the batch transmission scheduling process merges tasks according to the channel idle time window and the data block size;

[0011] During data transmission, the actual transmission delay of the time-critical task is continuously collected, and the channel reservation strategy is dynamically adjusted based on the deviation between the actual transmission delay and the preset delay threshold.

[0012] Furthermore, the data transmission task set is classified by latency sensitivity to generate a latency-critical task subset and a latency-tolerant task subset, including:

[0013] Extract the task generation timestamp and preset latest completion deadline for each transmission task from the data transmission task set;

[0014] Based on the task generation timestamp and the latest completion deadline, the maximum allowed transmission duration for each transmission task is calculated.

[0015] The maximum allowed transmission time is compared with a preset latency sensitivity threshold, which is determined based on the network reference round-trip time and jitter tolerance.

[0016] All transmission tasks whose maximum allowed transmission duration is less than the latency sensitivity threshold are classified as the latency-critical task subset, and the tasks in the latency-critical task subset have low tolerance for transmission interruption and delay.

[0017] All transmission tasks whose maximum allowed transmission duration is greater than or equal to the latency-sensitive threshold are classified into the latency-tolerant task subset, and tasks in the latency-tolerant task subset are allowed to be completed within a larger time window.

[0018] Furthermore, for the latency-critical subset of tasks, channel reservation processing based on predicted network state is performed, including:

[0019] Obtain a set of channel quality indicators for the current and historical moments, wherein the set of channel quality indicators includes reference signal received power, channel busy ratio, and bit error rate statistics;

[0020] Acquire real-time network load data, which includes the number of active user devices and the average length of the data queue.

[0021] The set of channel quality indicators and the real-time network load data are input into a pre-trained neural network model, and the neural network model outputs a predicted probability of channel availability for several future transmission time slots.

[0022] Based on the predicted channel availability probability, channels with availability probabilities consistently higher than the reservation threshold are selected to calculate channel reservation schemes for tasks in the latency-critical task subset. The channel reservation scheme specifies the starting time slot, duration time slot, and channel number reserved for each task.

[0023] The channel reservation scheme is distributed to relevant edge computing nodes and wireless access network devices to lock channel resources for a specified task in a specified time slot.

[0024] Further, based on the predicted channel availability probability, channels with availability probabilities consistently higher than the reservation threshold are selected to calculate channel reservation schemes for tasks in the latency-critical task subset, including:

[0025] Traverse all available channels and identify a set of candidate channels whose predicted channel availability probability is higher than the reserved threshold for N consecutive time slots in the future, where N is the minimum number of transmission time slots required based on the data size of the time-critical task and the theoretical channel rate.

[0026] For each task in the latency-critical task subset, select the lightest-loaded channel from the candidate channel set as the target reserved channel;

[0027] Calculate the actual number of transmission time slots required for each task based on the data volume of each task and the estimated effective transmission rate of the target reserved channel;

[0028] Starting from the earliest possible start time of the task's transmission, a number of consecutive actual transmission time slots are allocated on the target reserved channel to form an initial reservation plan;

[0029] Check if there are channel resource conflicts in the initial reservation plan. If there are conflicts, resolve them by fine-tuning the initial reservation time slot or switching to a channel with suboptimal load, and generate the final channel reservation scheme.

[0030] Further, for the latency-tolerant task subset, perform batch transmission scheduling processing based on dynamic aggregation of available channels, including:

[0031] Monitor the real-time status of all channels in the current network and record the predicted idle time window for each channel in the future;

[0032] For the tasks in the latency-tolerant task subset, cluster them according to their source node, target node and data size, and aggregate multiple tasks with the same source node and target node and small data size into a batch transmission task.

[0033] For each of the batch transmission tasks, one or more predicted idle time windows that are consecutive or adjacent in time are matched, such that the total length of the predicted idle time windows is sufficient to complete the transmission of the batch transmission task.

[0034] Based on the matching results, a batch transmission scheduling plan is generated, which specifies the start transmission time, the channel used, and the transmission order of each original subtask within the task for each batch transmission task.

[0035] When the predicted idle time window actually arrives, data transmission is activated according to the batch transmission scheduling plan, and channel quality is monitored during transmission. If the channel quality deteriorates to the switching threshold, a switch to a backup idle window or channel is triggered.

[0036] Furthermore, during data transmission, continuously collecting the actual transmission delay of the time-critical task, and dynamically adjusting the channel reservation strategy based on the deviation between the actual transmission delay and a preset delay threshold, includes:

[0037] After each time-critical task transmission is completed, its actual transmission delay metric is collected. The actual transmission delay metric is the total time from when the task leaves the source node to when it is acknowledged and received by the target node.

[0038] The actual transmission delay metric is compared with the preset delay threshold of the delay-critical task to calculate the delay deviation value;

[0039] If the delay deviation value is continuously positive and exceeds the tolerance limit, it is determined that the current channel reservation strategy is too aggressive, triggering a tightening adjustment of the reservation strategy. The tightening adjustment of the reservation strategy includes increasing the channel resource margin reserved for future tasks or selecting a channel with a higher availability probability prediction value.

[0040] If the delay deviation value is consistently negative or within an acceptable range, the current channel reservation strategy is determined to be moderate or conservative.

[0041] If the delay deviation value is negative for multiple consecutive monitoring periods, the reservation strategy relaxation adjustment is triggered. The reservation strategy relaxation adjustment includes appropriately reducing the reserved resource margin to improve spectrum utilization and allocating the released resources to the delay-tolerant task subset.

[0042] Furthermore, the tightening adjustment of the reservation strategy includes increasing the channel resource margin reserved for future tasks or selecting channels with higher predicted availability, including:

[0043] Based on historical delay deviation values, the number of additional protection time slots required for future predictions is calculated, and the number of additional protection time slots is proportional to the average delay deviation value.

[0044] When calculating the channel reservation scheme for new time-delay-critical tasks, the number of additional protection time slots is increased based on the minimum number of transmission time slots estimated according to the data volume and theoretical rate, and this number is used as the new total number of reserved time slots.

[0045] Update the channel selection criteria: when selecting a target reserved channel from the candidate channel set, prioritize channels whose predicted availability probability exceeds a higher-level reservation threshold.

[0046] When allocating reserved resources, a backup channel is reserved for the most critical tasks. The backup channel is activated when the quality of the primary channel deteriorates to the switching threshold.

[0047] Furthermore, when the predicted idle time window actually arrives, data transmission is activated according to the batch transmission scheduling plan, and channel quality is monitored during transmission. If the channel quality deteriorates to the handover threshold, a handover to a backup idle window or channel is triggered, including:

[0048] When the batch transmission task begins as scheduled, real-time channel quality monitoring is initiated, and the monitoring indicators include instantaneous bit error rate and signal-to-interference-to-noise ratio.

[0049] When the instantaneous bit error rate is continuously exceeded by the bit error rate threshold, or the signal-to-interference-plus-noise ratio is continuously lower than the signal-to-interference-plus-noise ratio threshold, it is determined that the current channel quality has deteriorated to the point of reaching the handover threshold.

[0050] Immediately suspend the current batch transmission task and query the backup idle time window or backup channel pre-allocated for the batch transmission task;

[0051] If a spare idle time window is available and time-coordinated, the remaining untransmitted data will be scheduled to continue transmission within the spare idle time window.

[0052] If no suitable alternative idle time window exists, but other idle channels are available, the remaining untransmitted data will be immediately switched to those other idle channels to continue transmission.

[0053] Furthermore, the method for constructing the pre-trained neural network model includes:

[0054] Obtain a historical network status dataset, which includes channel quality metrics and real-time network load data collected over multiple historical transmission cycles, as well as actual channel status labels for subsequent historical transmission cycles. The actual channel status labels are used to identify whether the channel is available in a specific future time slot.

[0055] The historical network state dataset is preprocessed, including filling in missing values, normalizing the data, and constructing model input samples according to a preset time series length.

[0056] A neural network model is constructed, which adopts an architecture combining a long short-term memory network and an attention mechanism. The long short-term memory network is used to extract time-dependent features from the network state data, and the attention mechanism is used to weight and focus on key time period information in the time-dependent features.

[0057] The preprocessed network state dataset is divided into a training set, a validation set, and a test set. The neural network model is trained using the training set. During the training process, the model hyperparameters are tuned using the validation set, and the model's prediction accuracy is evaluated using the test set.

[0058] Once the model's prediction accuracy reaches a preset threshold, the model parameters are saved, completing the construction of the pre-trained neural network model.

[0059] Furthermore, the method also includes an online learning step for the network state prediction model:

[0060] Collect the actual execution results of channel reservation processing and batch transmission scheduling processing, including the actual channels used, actual transmission delay, and actual interference events encountered;

[0061] The actual execution results are compared with the predicted network state on which the scheduling decision was based to generate a prediction error sample.

[0062] The pre-trained neural network model is incrementally trained using the prediction error samples to update the internal weight parameters of the neural network model;

[0063] Before the start of the next transmission cycle, a new channel availability probability prediction is made using the updated neural network model.

[0064] Compared with the prior art, the beneficial effects of the present invention are:

[0065] Based on preset latency thresholds and task generation timestamps, the data transmission task sets reported by multiple edge computing nodes within the current transmission cycle, which include source node identifiers, target node identifiers, data size, and task generation timestamps, are classified by latency sensitivity, dividing them into latency-critical task subsets and latency-tolerant task subsets. This classification method enables precise differentiation of tasks with different latency requirements, avoiding resource mismatches caused by neglecting task latency differences in conventional unified scheduling. It ensures that different types of tasks receive appropriate scheduling resources, reducing the possibility of latency-critical tasks being occupied by irrelevant tasks, while preventing latency-tolerant tasks from occupying critical channel resources, thus achieving precise matching between tasks and scheduling resources.

[0066] For the latency-critical task subset, channel reservation is performed based on predicted network conditions derived from historical channel quality metrics and real-time network load data. For the latency-tolerant task subset, tasks are merged according to channel idle time windows and data block sizes, and batch transmission scheduling is performed based on dynamic aggregation of available channels. Channel reservation by predicting network conditions can avoid network congestion periods and low-quality channels in advance, reducing latency fluctuations during transmission and avoiding the resource waste of conventional static channel reservation and the latency instability of random channel allocation. Batch transmission through dynamic aggregation of available channels can fully utilize idle channel resources, rationally merge multiple latency-tolerant tasks, avoid resource consumption caused by frequent channel occupation by individual small-data-volume tasks, improve channel resource utilization, and reduce the overall transmission overhead of latency-tolerant tasks. Attached Figure Description

[0067] Figure 1 This is a flowchart illustrating the steps of the low-latency data transmission scheduling method for 5G edge computing described in this invention.

[0068] Figure 2 A flowchart for classifying latency sensitivity;

[0069] Figure 3 A flowchart for channel reservation processing. Detailed Implementation

[0070] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.

[0071] See Figure 1 This invention provides a low-latency data transmission scheduling method for 5G edge computing, the specific method including:

[0072] The system acquires a set of data transmission tasks reported by multiple edge computing nodes within the current transmission cycle. This set includes the source node identifier, target node identifier, data size, and task generation timestamp for each task. The data transmission task set is then categorized by latency sensitivity, generating latency-critical and latency-tolerant task subsets. The latency sensitivity classification is calculated based on a preset latency threshold and the task generation timestamp. For the latency-critical task subset, channel reservation processing based on predicted network status is performed. The predicted network status is derived from historical channel quality indicators and real-time network load data. For the latency-tolerant task subset, batch transmission scheduling processing based on dynamic aggregation of available channels is performed. This batch transmission scheduling merges tasks according to the channel idle time window and data block size. During data transmission, the actual transmission latency of latency-critical tasks is continuously collected, and the channel reservation strategy is dynamically adjusted based on the deviation between the actual transmission latency and the preset latency threshold.

[0073] In one embodiment of the present invention, the data transmission task set is classified by latency sensitivity to generate a latency-critical task subset and a latency-tolerant task subset. The specific process is as follows: (See [reference]) Figure 2 The process extracts the task generation timestamp and preset latest completion deadline for each transmission task from the data transmission task set. Based on these timestamps and deadlines, the maximum allowable transmission duration for each task is calculated. This maximum allowable transmission duration is then compared to a preset latency sensitivity threshold, determined based on network baseline round-trip time and jitter tolerance. All transmission tasks with a maximum allowable transmission duration less than the latency sensitivity threshold are categorized as a latency-critical subset, exhibiting low tolerance for transmission interruptions and delays. Conversely, all transmission tasks with a maximum allowable transmission duration greater than or equal to the latency sensitivity threshold are categorized as a latency-tolerant subset, allowing completion within a larger time window.

[0074] In practical implementation, the execution process relies on a centralized scheduler in the 5G edge computing network. At the beginning of each transmission cycle, the centralized scheduler collects a set of data transmission tasks reported by multiple edge computing nodes. In this implementation, each transmission task in the data transmission task set is explicitly defined and formatted, including the source node identifier, target node identifier, data size, and task generation timestamp. The task generation timestamp records the precise system time when the task was generated at the source node. To perform latency sensitivity classification, the centralized scheduler also needs to obtain the preset latest completion deadline for each transmission task. This deadline is predefined by the application or quality of service policy that generated the transmission task and is stored along with the task generation timestamp in the task metadata.

[0075] In some embodiments, the centralized scheduler sequentially extracts the task generation timestamp and preset latest completion deadline for each transmission task from the data transmission task set. It can be understood that the time difference between the task generation timestamp and the latest completion deadline defines the total available time from the generation of the transmission task to its mandatory completion. Based on this, the centralized scheduler calculates the maximum allowed transmission duration for each transmission task. The maximum allowed transmission duration is calculated by subtracting the task generation timestamp from the latest completion deadline. In an example scenario, the data transmission task set contains three tasks: Task 1 is generated at timestamp 100 milliseconds, with a latest completion deadline of 130 milliseconds, so the maximum allowed transmission duration is 30 milliseconds; Task 2 is generated at timestamp 105 milliseconds, with a latest completion deadline of 125 milliseconds, so the maximum allowed transmission duration is 20 milliseconds; and Task 3 is generated at timestamp 110 milliseconds, with a latest completion deadline of 200 milliseconds, so the maximum allowed transmission duration is 90 milliseconds. These calculated values ​​are the direct basis for subsequent classification.

[0076] In practical implementation, the preset latency sensitivity threshold is a key threshold value used for decision-making. The determination of the latency sensitivity threshold depends on the network baseline round-trip time (RTT) and jitter tolerance. The network baseline RTT is obtained by repeatedly measuring the statistical minimum of end-to-end latency under network idle or light load conditions. The jitter tolerance is set based on the statistical variance of historical data transmission latency or the acceptable latency fluctuation range for the application layer. A feasible calculation method is expressed by the following formula:

[0077]

[0078] in: Indicates the latency-sensitive threshold. Indicates the network baseline round-trip time. This indicates the jitter tolerance. For example, if the measured network baseline round-trip latency is 5 milliseconds and the jitter tolerance set according to the policy is 3 milliseconds, the latency sensitivity threshold is set to 8 milliseconds. It can be understood that the latency sensitivity threshold is a global or regional configuration parameter used to distinguish between task types with strict latency requirements and those with relatively lenient requirements.

[0079] In practice, the classification operation is implemented by comparing the maximum allowed transmission duration of each transmission task with a latency-sensitive threshold. The centralized scheduler iterates through all calculated maximum allowed transmission duration values ​​and compares each value with the latency-sensitive threshold. For all transmission tasks whose maximum allowed transmission duration is less than the latency-sensitive threshold, the centralized scheduler classifies them into a latency-critical task subset. In the previous example, Task 1's maximum allowed transmission duration of 30 milliseconds is greater than the latency-sensitive threshold of 8 milliseconds, Task 2's maximum allowed transmission duration of 20 milliseconds is greater than the latency-sensitive threshold of 8 milliseconds, and Task 3's maximum allowed transmission duration of 90 milliseconds is greater than the latency-sensitive threshold of 8 milliseconds. If the latency-sensitive threshold is 8 milliseconds, then none of the three tasks in the example belong to the latency-critical task subset. If there is a Task 4 with a maximum allowed transmission duration of 6 milliseconds, since 6 milliseconds is less than the latency-sensitive threshold of 8 milliseconds, Task 4 will be classified as a latency-critical task. Tasks in the latency-critical task subset have low tolerance for transmission interruptions and delays, and their scheduling must be prioritized. For all transmission tasks whose maximum allowed transmission duration is greater than or equal to the latency-sensitive threshold, the centralized scheduler categorizes them into a latency-tolerant task subset. In the example scenario, Task 1, Task 2, and Task 3 all have a maximum allowed transmission duration greater than 8 milliseconds, therefore they are categorized into the latency-tolerant task subset. Tasks in the latency-tolerant task subset are allowed to complete within a larger time window, and their scheduling can more flexibly utilize idle network resources. Optionally, the two task subsets generated after classification will serve as explicit inputs for subsequent differentiated scheduling processing.

[0080] In some embodiments, the latency-sensitive threshold can be dynamically fine-tuned based on network load conditions. For example, during periods of light network load, the baseline round-trip time may be lower, and the latency-sensitive threshold can be lowered accordingly to achieve finer differentiation; during periods of network congestion, jitter tolerance may need to be moderately relaxed, and the latency-sensitive threshold can be raised accordingly to avoid marking too many tasks as critical and causing resource contention. Optionally, this dynamic adjustment is automatically performed based on statistical values ​​of network performance metrics collected over historical periods.

[0081] In one embodiment of the present invention, for a subset of delay-critical tasks, channel reservation processing based on predicted network state is performed. The process is as follows: (See [link to relevant documentation]). Figure 3The process involves acquiring a set of channel quality metrics for the current and historical moments, including reference signal received power, channel busy ratio, and bit error rate statistics. Simultaneously, real-time network load data, including the number of active user devices and the average data queue length, is also obtained. The channel quality metrics set and real-time network load data are input into a pre-trained neural network model, which outputs a predicted channel availability probability for several future transmission time slots. Based on the predicted channel availability probability, channels with availability probabilities consistently higher than a reservation threshold are selected to calculate channel reservation schemes for tasks within the latency-critical task subset. When calculating the channel reservation scheme, all available channels are first traversed to identify a set of candidate channels whose predicted availability probability is higher than the reservation threshold for N consecutive time slots, where N is the minimum number of transmission time slots estimated based on the data volume of the latency-critical task and the theoretical channel rate. Then, for each task in the latency-critical task subset, the lightest-loaded channel is selected from the candidate channel set as the target reserved channel. Based on the data volume of each task and the estimated effective transmission rate of the target reserved channel, the actual number of transmission time slots required for each task is calculated. Starting with the earliest possible start time for transmission of the task, a number of consecutive actual transmission time slots are allocated on the target reserved channel to form an initial reservation plan. The initial reservation plan is checked for channel resource conflicts. If a conflict exists, the initial reservation time slot is fine-tuned or the channel with the next lowest load is switched for resolution, generating the final channel reservation scheme. The final channel reservation scheme is then distributed to the relevant edge computing nodes and wireless access network devices to lock channel resources for the specified task in the designated time slot.

[0082] In practice, after obtaining a subset of latency-critical tasks, the centralized scheduler initiates a channel reservation process based on predicted network conditions. Specifically, the centralized scheduler obtains a set of channel quality indicators (CMIs) for the current and historical moments from network probes and base station controllers. These CMIs include reference signal received power, channel busy ratio, and bit error rate (BER) statistics. Reference signal received power reflects signal strength, the channel busy ratio indicates the proportion of time the channel is occupied, and the BER statistics describe the reliability of data transmission. Simultaneously, the centralized scheduler obtains real-time network load data from core network elements and access network equipment. This real-time network load data includes the number of active user devices and the average data queue length. The number of active user devices characterizes the current terminal scale of the access network, and the average data queue length reflects the congestion level of data awaiting forwarding.

[0083] In some embodiments, the centralized scheduler organizes the acquired set of channel quality indicators and real-time network load data into a time-series format to form a model input vector. It can be understood that the set of channel quality indicators and real-time network load data together constitute a quantitative description of the current and historical network state. This input vector is fed into a pre-trained neural network model, which is deployed within the centralized scheduler. Based on the aforementioned input vector, the pre-trained neural network model analyzes its temporal patterns and correlations, and outputs a predicted value for the availability of each potential channel within several consecutive transmission time slots in the future. This predicted value is a probability between 0 and 1, called the channel availability probability prediction value.

[0084] In practice, based on the predicted channel availability probability output by the pre-trained neural network model, the centralized scheduler performs channel selection and reservation scheme calculations. The centralized scheduler sets a fixed reservation threshold, for example, 0.85. It iterates through all available logical channels, identifying those whose predicted availability probability is higher than the reservation threshold of 0.85 for the next N consecutive time slots; these channels form the candidate channel set. The parameter N is the minimum number of transmission time slots estimated based on the data size of the specific task within the latency-critical task subset and the theoretical transmission rate of the channel. In an example scenario, if a latency-critical task has 10 megabits of data and the estimated effective transmission rate of the selected channel is 5 megabits per time slot, then at least two consecutive time slots are needed to complete this task, and N is set to 2. The centralized scheduler executes subsequent steps independently for each task within the latency-critical task subset.

[0085] For each task in the latency-critical task subset, the centralized scheduler selects a target reserved channel from the candidate channel set. The selection criterion is the lightest load, which can be measured based on a weighted average of the channel's historical busy ratios or the number of active user equipment associated with that channel. For example, if the candidate channel set includes channel A and channel B, with channel A having a historical average busy ratio of 30% and channel B having a historical average busy ratio of 50%, the centralized scheduler selects the lighter-loaded channel A as the target reserved channel for the current task. After selecting the target reserved channel, the centralized scheduler calculates the actual number of transmission slots required for the task based on the data volume of the task and the estimated effective transmission rate of the target reserved channel. The estimated effective transmission rate is estimated by multiplying the channel's theoretical rate by the predicted channel availability probability. The calculation formula can be expressed as:

[0086]

[0087] in: Indicates the actual number of transmission time slots required for the task. This indicates the size of the task's data. This represents the estimated effective transmission rate of the target reserved channel. This indicates the floor function (rounding up). For example, the amount of data in a task. The target channel is estimated to have an effective transmission rate of 7 megabits per second. If the value is 3 megabits per time slot, then the calculation is as follows: Each time slot.

[0088] Understandably, after calculating the actual number of transmission time slots, the centralized scheduler uses the earliest possible start time of a task as the starting point and attempts to allocate a consecutive number of actual transmission time slots on the target reserved channel, thus forming an initial reservation plan. The earliest possible start time of a task is usually equal to or slightly later than the task generation timestamp, taking into account necessary processing delays. The centralized scheduler maintains a global channel time slot resource mapping table in memory to record the reservation status of all channels in different time slots. When attempting to generate initial reservation plans for multiple tasks, resource conflicts may occur where the same time slot on the same channel is allocated to different tasks. For example, task X is planned to be transmitted in time slots 5 to 7 of channel A, while the initial reservation plan for task Y also occupies time slot 6 of channel A, resulting in a conflict. In specific implementations, the centralized scheduler checks whether such channel resource conflicts exist in all initial reservation plans. If a conflict exists, the centralized scheduler initiates a conflict resolution mechanism. Optionally, conflict resolution can be achieved by fine-tuning the reserved start time slots of one or more conflicting tasks, for example, postponing the start time slot of task Y to time slot 8. Optionally, conflict resolution can also be achieved by switching the conflicting task to another channel in the candidate channel set with the second-lowest load, for example, switching the target reserved channel for task Y from channel A to channel B with a slightly higher load. Through a combination of one or more of the above methods, the centralized scheduler generates the final conflict-free channel reservation scheme. The final channel reservation scheme explicitly specifies the starting time slot, duration time slot, and channel number reserved for each task in the latency-critical task subset. The centralized scheduler distributes this channel reservation scheme to the relevant source nodes, the edge computing nodes where the target nodes are located, and the radio access network equipment responsible for radio resource scheduling. After receiving the scheme, the radio access network equipment will lock the channel resources for the specified task in the specified time slot, prohibiting other data transmission requests from occupying them, thereby ensuring that the transmission of latency-critical tasks can proceed exclusively as planned.

[0089] In one embodiment of the present invention, for a subset of delay-tolerant tasks, a batch transmission scheduling process based on dynamic aggregation of available channels is performed. The process involves: monitoring the real-time status of all channels in the current network and recording the predicted idle time window for each channel within a future period. Tasks in the delay-tolerant task subset are clustered based on their source node, target node, and data size, aggregating multiple tasks with the same source and target nodes and small data sizes into a single batch transmission task. For each batch transmission task, one or more temporally consecutive or adjacent predicted idle time windows are matched, ensuring the total length of the predicted idle time windows is sufficient to complete the transmission of the batch transmission task. Based on the matching results, a batch transmission scheduling plan is generated, specifying the start transmission time, the channel used, and the transmission order of the original subtasks within each batch transmission task. When the predicted idle time window actually arrives, data transmission is activated according to the batch transmission scheduling plan, and channel quality is monitored during transmission. If the channel quality deteriorates to the switching threshold, a switch to a backup idle window or channel is triggered. The handover process is as follows: When a batch transmission task begins as planned, real-time channel quality monitoring is initiated, with monitoring indicators including instantaneous bit error rate (BER) and signal-to-interference-plus-noise ratio (SNR). When the instantaneous BER continuously exceeds the BER threshold, or the SNR remains below the SNR threshold, the current channel quality is deemed to have deteriorated to the handover threshold. The current batch transmission task is immediately suspended, and a pre-allocated spare idle time window or spare channel is queried. If a usable and time-coordinated spare idle time window exists, the remaining untransmitted data is scheduled to continue transmission within that window. If no suitable spare idle time window exists, but other idle channels are available, the remaining untransmitted data is immediately switched to those channels for transmission.

[0090] In practical implementation, after obtaining a subset of delay-tolerant tasks, the centralized scheduler initiates batch transmission scheduling processing based on dynamic aggregation of available channels. The centralized scheduler continuously monitors the real-time status of all logical channels in the current 5G network, with real-time status information obtained periodically from base stations and network probes. Based on historical channel occupancy statistics and real-time load prediction algorithms, the centralized scheduler records the predicted idle time window for each channel within a future period. The predicted idle time window, represented by start and end time slots, indicates the time period during which the channel is not occupied by high-priority services. For clarity, an exemplary predicted idle time window record is shown in Table 1.

[0091] Table 1: Predicted Channel Idle Time Window Table

[0092] In practical implementation, the centralized scheduler performs clustering operations on tasks within the latency-tolerant task subset. Clustering is based on three attributes: source node identifier, target node identifier, and data size. The centralized scheduler aggregates multiple independent tasks with identical source and target node identifiers, and whose data size is less than a preset aggregation size threshold, into a logical batch transmission task. In an example scenario, the latency-tolerant task subset includes tasks A, B, and C. Task A has edge node X as its source node, edge node Y as its target node, and a data size of 2 megabits; Task B has edge node X as its source node, edge node Y as its target node, and a data size of 1.5 megabits; Task C has edge node Z as its source node, edge node Y as its target node, and a data size of 3 megabits. Assuming the aggregation size threshold is 5 megabits, tasks A and B, because they share the same source node X and target node Y, and their respective data sizes (2 megabits and 1.5 megabits, respectively) are both less than 5 megabits, are aggregated into a single batch transmission task, denoted as batch task AB. Task C is not included in this batch task because its source node is different. The total data volume of the batch transfer task is the sum of the data volumes of all its subtasks, calculated using the following formula:

[0093]

[0094] in: This represents the total data volume of the aggregated batch transfer task. This indicates the number of subtasks that are aggregated. Indicates the first The size of the data in each subtask. In the example, the total data size of batch tasks A and B. Megabits.

[0095] In some embodiments, a predicted idle time window is matched for each aggregated batch transmission task. The centralized scheduler traverses the predicted channel idle time window table to find one or more temporally consecutive or adjacent predicted idle time windows for the batch transmission task. The goal of the matching is to ensure that the total length (in timeslots) of the selected predicted idle time windows is sufficient to complete the transmission of the batch transmission task. The total number of timeslots required is determined by the total data volume of the batch transmission task. This is obtained by dividing by the estimated available transmission rate of the matched channel. For example, to transmit batch task AB (3.5 Mbps), assuming matched channel 1 has an estimated available transmission rate of 1 Mbps / slot, at least 4 consecutive time slots are required. The centralized scheduler finds in the predicted idle time window table that channel 1 has a predicted idle time window of length 10 in time slots 10 to 19, which meets the requirement, and therefore matches this window with batch task AB. Understandably, if the length of a single predicted idle time window is insufficient, the centralized scheduler will attempt to combine predicted idle time windows that are time-adjacent on the same channel or on multiple different channels, but cross-channel transmission may involve handover overhead.

[0096] In practice, based on the matching results, the centralized scheduler generates a batch transmission scheduling plan. The batch transmission scheduling plan is a structured list that explicitly specifies the starting transmission time slot, the channel number used, and the transmission order of each original subtask within the task for each batch transmission task. The transmission order can be arranged according to the data size of the subtasks or the generation timestamp. For example, the scheduling plan entry generated for batch task AB is: starting time slot 10, using channel 1, transmission order is [task A, task B]. The centralized scheduler distributes this plan to the relevant edge computing nodes.

[0097] In practice, when the predicted idle time window actually arrives, the relevant edge computing nodes activate data transmission according to the batch transmission scheduling plan. Simultaneously with transmission initiation, a real-time channel quality monitoring process is started. This monitoring process continuously collects physical layer metrics of the currently used channel, with core metrics including instantaneous bit error rate (BER) and signal-to-interference-plus-noise ratio (SINR). The BER is calculated after demodulation at the receiver, and the SINR is obtained by measuring the ratio of the reference signal strength to the interference noise power. The centralized scheduler or receiving nodes set specific threshold values, such as BER thresholds and SINR thresholds. When the BER continuously exceeds the BER threshold or the SINR continuously falls below the SINR threshold, the system determines that the current channel quality has deteriorated to the preset handover threshold. The setting of the handover threshold requires a trade-off between handover overhead and transmission reliability.

[0098] In some embodiments, once a handover determination is triggered, the system immediately executes the handover procedure. It is understood that the primary action of the handover is to immediately suspend the transmission of the current batch transmission task. Subsequently, the system queries whether a spare idle time window or a spare channel has been allocated for the batch transmission task when the batch transmission scheduling plan was generated. For example, batch task AB uses time slots 10-19 of channel 1 in the initial plan, but may have pre-allocated time slots 50-59 of channel 2 as a spare idle time window. If such an available and temporally contiguous spare idle time window exists, the system reschedules the remaining untransmitted data to the spare idle time window for continued transmission, which may result in a delay in transmission completion time. Optionally, if no suitable spare idle time window exists, but other channels are detected to be idle, the system may perform an immediate handover. For example, if the quality of channel 1 deteriorates and channel 3 is detected to be currently idle, the system immediately switches the remaining untransmitted data to channel 3 for continued transmission. Optionally, after switching to a new channel, the system reassesses the transmission time of the remaining data and may update subsequent scheduling plans. The entire handover process aims to leverage the flexibility offered by dynamic channel aggregation to improve the overall utilization of spectrum resources while ensuring the eventual completion of delay-tolerant tasks.

[0099] In one embodiment of the present invention, during data transmission, the actual transmission delay of latency-critical tasks is continuously collected, and the channel reservation strategy is dynamically adjusted based on the deviation between the actual transmission delay and a preset latency threshold. Specifically, the adjustment process is as follows: After each latency-critical task is completed, its actual transmission delay metric is collected. This metric is the total time from when the task leaves the source node to when it is acknowledged by the target node. The actual transmission delay metric is compared with the preset latency threshold for that latency-critical task, and a latency deviation value is calculated. If the latency deviation value is consistently positive and exceeds the tolerance limit, the current channel reservation strategy is deemed too aggressive, triggering a tightening adjustment of the reservation strategy. If the latency deviation value is negative for multiple consecutive monitoring periods, a relaxing adjustment of the reservation strategy is triggered. Tightening the reservation strategy includes increasing the channel resource margin reserved for future tasks or selecting channels with higher predicted availability. The specific operation of the tightening adjustment is: based on historically statistically analyzed latency deviation values, the number of additional protection slots required for future predictions is calculated. This number of additional protection slots is proportional to the average latency deviation value. When calculating channel reservation schemes for new latency-critical tasks, the number of additional protection time slots is added to the minimum number of transmission time slots estimated based on data volume and theoretical rate, resulting in a new total number of reserved time slots. The channel selection criteria are updated; when selecting target reserved channels from the candidate channel set, channels with predicted availability probabilities exceeding higher-level reservation thresholds are prioritized. During reserved resource allocation, backup channels are reserved for the highest-criticality tasks; these backup channels are activated when the quality of the primary channel deteriorates to the handover threshold. Relaxation adjustments to the reservation strategy include moderately reducing reserved resource margins to improve spectrum utilization and allocating the released resources to a subset of latency-tolerant tasks.

[0100] In practice, the centralized scheduler continuously monitors latency and dynamically adjusts strategies during data transmission for latency-critical tasks based on the channel reservation scheme. After each latency-critical task is completed, the relevant target node or intermediate network element reports the actual transmission latency metric to the centralized scheduler. The actual transmission latency metric is defined as the total time from the moment the task data leaves the source node to the moment the task data is successfully received and acknowledged by the target node; this time includes transmission latency, queuing latency, and potential processing latency. The centralized scheduler maintains a latency record database to store the task identifier, preset latency threshold, and reported actual transmission latency metric for each completed latency-critical task.

[0101] In some embodiments, after receiving the actual transmission delay measurement, the centralized scheduler initiates a delay deviation calculation process. The centralized scheduler extracts the preset delay threshold for the corresponding delay-critical task from the database. The preset delay threshold is the maximum allowable transmission duration upon which the task relies during the delay sensitivity classification phase. It can be understood that the purpose of calculating the delay deviation value is to quantify the gap between actual performance and preset requirements. The formula for calculating the delay deviation value is as follows:

[0102]

[0103] in: Indicates task The time delay deviation value, Indicates task The actual transmission delay metric Indicates task The preset delay threshold. If This indicates that the actual transmission delay exceeds the preset threshold; if This indicates that the actual transmission latency meets or is better than the preset threshold. For ease of analysis, the centralized scheduler records a sequence of latency deviation values ​​over several recent consecutive monitoring periods (e.g., the last 10 tasks). See Table 2 for an example of a latency deviation record:

[0104] Table 2: Time Delay Deviation Monitoring Record Table

[0105] In practice, the centralized scheduler makes policy adjustment decisions based on latency deviation records. The centralized scheduler sets a positive tolerance upper limit, for example, +1.0 milliseconds. It analyzes consecutive latency deviation values. If the latency deviation value is consistently positive and exceeds the tolerance upper limit (+1.0 milliseconds), it determines that the current channel reservation policy is too aggressive, and the reserved resource margin is insufficient to cope with network fluctuations, thus triggering a tightening adjustment of the reservation policy. For example, as shown in the table above, the latency deviation values ​​of tasks_101,_103, and_104 all exceed +1.0 milliseconds, indicating that the reservation policy may be too aggressive. Conversely, if the latency deviation value is negative for several consecutive monitoring periods, it indicates that the reservation policy may be too conservative, resulting in wasted spectrum resources, thus triggering a relaxation adjustment of the reservation policy. If the latency deviation value is within an acceptable range, it determines that the current channel reservation policy is appropriate and can remain unchanged.

[0106] Tightening the reservation strategy involves increasing the channel resource margin reserved for future missions or selecting channels with higher predicted availability probabilities. In practice, increasing the resource margin is achieved by calculating and adding additional protection time slots. The centralized scheduler calculates the number of additional protection time slots required for future predictions based on historically statistically analyzed delay deviations. The number of additional protection time slots is proportional to the historical average delay deviation. One calculation method is:

[0107]

[0108] in: This indicates the number of additional protection time slots required. This represents the arithmetic mean of all positive time delay deviations within the monitoring period deemed "too aggressive." The standard duration of a transmission time slot is represented. This indicates the rounding up sign. For example, if the average time delay deviation value... +1.5 milliseconds, duration of a single time slot If it is 1 millisecond, then An additional protection time slot. When calculating channel reservation schemes for new delay-critical tasks, the centralized scheduler will add this to the minimum number of transmission time slots estimated based on the data volume and theoretical rate. The number of additional protection time slots is used as the total number of new reserved time slots.

[0109] It is understandable that, in addition to increasing time slots, tightening adjustments also include updating channel selection criteria. In some embodiments, the centralized scheduler sets a higher reservation threshold, for example, increasing it from 0.85 to 0.95. When selecting a target reserved channel from the candidate channel set for a task, channels with a predicted availability probability exceeding 0.95 are prioritized, instead of the original 0.85. Optionally, during resource allocation, the centralized scheduler also reserves a backup channel for the task with the highest criticality level in the task subset (such as the task with the lowest latency threshold). The backup channel is another channel with a similarly high predicted availability probability. When real-time quality monitoring of the primary channel detects signal quality deterioration to a preset switching threshold, the system immediately activates the backup channel for transmission to provide additional reliability assurance.

[0110] The reservation strategy relaxation adjustment is triggered when negative latency deviation values ​​are continuously detected. The relaxation adjustment includes moderately reducing the resource margin reserved for tasks. For example, the number of additional protection slots previously added for tightening adjustments can be reduced. This can even be done by setting it to zero. Understandably, the direct purpose of relaxing the adjustment is to improve spectrum utilization. By reducing the conservative reservation for latency-critical tasks, some channel resources can be freed up. The centralized scheduler dynamically allocates these freed-up channel resources to a subset of latency-tolerant tasks for their batch transmission scheduling, thereby improving the overall utilization efficiency of network resources. Optionally, the channel selection criteria can also be relaxed simultaneously, for example, by lowering the reservation threshold from 0.95 to 0.85, allowing more channels to qualify for the candidate channel set, thus increasing scheduling flexibility.

[0111] In one embodiment of the present invention, the pre-trained neural network model is constructed as follows: A historical network state dataset is acquired, which includes channel quality indicators and real-time network load data collected over multiple historical transmission cycles, as well as actual channel state labels for subsequent historical transmission cycles. These actual channel state labels are used to identify whether the channel is available in a specific future time slot. The historical network state dataset is preprocessed, including filling in missing values, normalizing the data, and constructing model input samples according to a preset time series length. A neural network model is constructed, employing an architecture combining a Long Short-Term Memory (LSTM) network and an attention mechanism. The LSM network is used to extract time-dependent features from the network state data, and the attention mechanism is used to weight and focus on key time-period information within the time-dependent features. The preprocessed network state dataset is divided into a training set, a validation set, and a test set. The neural network model is trained using the training set, and the model's hyperparameters are tuned using the validation set during training. The model's prediction accuracy is evaluated using the test set. Once the model's prediction accuracy reaches a preset threshold, the model parameters are saved, completing the construction of the pre-trained neural network model. The method also includes an online learning step for the network state prediction model: collecting actual execution results during channel reservation processing and batch transmission scheduling, including the actual channels used, actual transmission delays, and actual interference events encountered. The actual execution results are compared with the predicted network state used when making previous scheduling decisions to generate prediction error samples. These prediction error samples are then used to incrementally train a pre-trained neural network model, updating its internal weight parameters. Before the start of the next transmission cycle, the updated neural network model is used to predict the new channel availability probability.

[0112] In practice, building a pre-trained neural network model begins with data preparation. A centralized scheduler or a dedicated data acquisition unit obtains historical network state datasets. These datasets contain channel quality metrics and real-time network load data collected over multiple historical transmission periods. Channel quality metrics include reference signal received power, channel busy ratio, and bit error rate statistics. Real-time network load data includes the number of active user devices and the average data queue length. The historical network state datasets also contain actual channel state labels corresponding to subsequent moments in these historical periods. Each actual channel state label is a binary flag indicating whether a specific channel is actually available for data transmission in a specific future time slot.

[0113] In practice, preprocessing the historical network state dataset is a crucial step in model building. Preprocessing includes imputing missing values ​​in the data, using methods such as forward imputation or linear interpolation. Preprocessing also includes data normalization, mapping channel quality indicators and real-time network load data of different dimensions and ranges to a unified numerical interval, such as [0,1]. The final step in preprocessing is to construct model input samples according to a preset time series length. Each input sample is a sequence with a time step of T, where each time point corresponds to a historical network state feature vector, and the output label for that sample is the actual channel state label for a specific future time slot. For example, a sample could use network state data from the past 10 time slots (T=10) as input, with the actual channel availability of the third time slot after the current time slot as the label.

[0114] In some embodiments, the neural network model employs an architecture combining a Long Short-Term Memory (LSTM) network and an attention mechanism. The LTM layer extracts time-dependent features from the network state data, capturing dynamic patterns of channel quality and network load over time. The attention mechanism layer receives features from all time steps output by the LTM layer and weights and focuses on key time periods within these time-dependent features, allowing the model to pay more attention to historical periods that have the greatest impact on predicting future channel states. It can be understood that the combination of LTM and the attention mechanism enables the model to more effectively learn valid representations for prediction from historical sequences. The final output layer of this neural network model is typically a fully connected layer with a sigmoid activation function, outputting a scalar between 0 and 1, representing the probability that the predicted channel will be available in the future target time slot.

[0115] In practice, after preprocessing and model architecture definition, the preprocessed network state dataset is divided into three mutually exclusive subsets: training set, validation set, and test set. The neural network model is trained using the training set, with backpropagation and an optimizer used to minimize prediction error during training. The loss function is typically a binary cross-entropy. An example loss function is calculated as follows:

[0116]

[0117] in: Indicates the loss value. This indicates the number of samples in a training batch. Indicates the first The actual channel state label (0 or 1) of each sample. Indicates the neural network model for the first... The predicted output probability for each sample. During training, the model's hyperparameters, including the learning rate and the number of hidden units in the Long Short-Term Memory network, are tuned using a validation set to prevent overfitting on the training set. After training, the model's prediction accuracy is evaluated using a test set; prediction accuracy is the proportion of times the model makes correct predictions on unseen test data. When the model's prediction accuracy reaches a preset threshold, such as 95%, the central scheduler saves the model's structure and internal weight parameters, completing the construction of the pre-trained neural network model and deploying it for online prediction.

[0118] In some embodiments, the method further includes an online learning step for the network state prediction model, enabling the pre-trained neural network model to adapt to changes in the network environment. During the online learning process, the centralized scheduler continuously collects the actual execution results of the channel reservation processing and batch transmission scheduling processing. The collected actual execution results include the actual channel number used, the actual transmission delay of the task, and the type and occurrence time of the interference events actually encountered during transmission. The centralized scheduler compares the actual execution results with the predicted network state on which the scheduling decision was previously made, generating a prediction error sample. For example, if the model predicted that the probability of channel A being available in time slot t was 0.9, but the channel failed to be used due to sudden interference during actual transmission, a negative sample (high predicted probability, actually unavailable) is generated, where the true label is 0, and the model's original prediction value is 0.9.

[0119] In practice, the generated prediction error samples are used to incrementally train the deployed pre-trained neural network model. Incremental training does not start from scratch; instead, it updates the internal weight parameters of the neural network model with a small learning rate using newly collected prediction error samples, building upon the existing model's weight parameters. This online learning approach allows the model to absorb new network state patterns and gradually correct prediction biases. Optionally, online learning can be performed periodically or triggered after prediction errors accumulate to a certain amount. After online learning is complete, before the start of the next transmission cycle, the centralized scheduler uses the updated neural network model to predict new channel availability probabilities, thus forming a closed-loop model self-optimization process. This ensures that the channel state prediction capability continuously adapts to actual network changes over time.

[0120] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A low-latency data transmission scheduling method for 5G edge computing, characterized in that, Includes the following steps: Obtain the set of data transmission tasks reported by multiple edge computing nodes within the current transmission cycle. The set of data transmission tasks includes the source node identifier, target node identifier, data size, and task generation timestamp for each transmission task. The data transmission task set is classified by latency sensitivity to generate a latency-critical task subset and a latency-tolerant task subset. The latency sensitivity classification is calculated based on a preset latency threshold and the task generation timestamp. For the latency-critical subset of tasks, channel reservation processing based on predicted network state is performed, wherein the predicted network state is derived from historical channel quality indicators and real-time network load data; For the latency-tolerant subset of tasks, a batch transmission scheduling process based on dynamic aggregation of available channels is executed, wherein the batch transmission scheduling process merges tasks according to the channel idle time window and the data block size; During data transmission, the actual transmission delay of the time-critical task is continuously collected, and the channel reservation strategy is dynamically adjusted based on the deviation between the actual transmission delay and a preset delay threshold, including: After each time-critical task transmission is completed, its actual transmission delay metric is collected. The actual transmission delay metric is the total time from when the task leaves the source node to when it is acknowledged and received by the target node. The actual transmission delay metric is compared with the preset delay threshold of the delay-critical task to calculate the delay deviation value; If the delay deviation value is continuously positive and exceeds the tolerance limit, it is determined that the current channel reservation strategy is too aggressive, triggering a tightening adjustment of the reservation strategy. The tightening adjustment of the reservation strategy includes increasing the channel resource margin reserved for future tasks or selecting a channel with a higher availability probability prediction value. If the delay deviation value is consistently negative or within an acceptable range, the current channel reservation strategy is determined to be moderate or conservative. If the delay deviation value is negative for multiple consecutive monitoring periods, the reservation strategy relaxation adjustment is triggered. The reservation strategy relaxation adjustment includes appropriately reducing the reserved resource margin to improve spectrum utilization and allocating the released resources to the delay-tolerant task subset. Based on the predicted channel availability probability, channels with availability probabilities consistently higher than the reservation threshold are selected. Channel reservation schemes are then calculated for tasks within the latency-critical task subset, including: Traverse all available channels and identify a set of candidate channels whose predicted channel availability probability is higher than the reserved threshold for N consecutive time slots in the future, where N is the minimum number of transmission time slots required based on the data size of the time-critical task and the theoretical channel rate. For each task in the latency-critical task subset, select the lightest-loaded channel from the candidate channel set as the target reserved channel; Calculate the actual number of transmission time slots required for each task based on the data volume of each task and the estimated effective transmission rate of the target reserved channel; Starting from the earliest possible start time of the task's transmission, a number of consecutive actual transmission time slots are allocated on the target reserved channel to form an initial reservation plan; Check if there are channel resource conflicts in the initial reservation plan. If there are conflicts, resolve them by fine-tuning the initial reservation time slot or switching to a channel with the second-best load, and generate the final channel reservation scheme. The estimated effective transmission rate is calculated by multiplying the theoretical channel rate by the predicted channel availability probability, as expressed by the following formula: ； in, Indicates the actual number of transmission time slots required for the task. This indicates the size of the task's data. This represents the estimated effective transmission rate of the target reserved channel. The symbol indicates rounding up.

2. The low-latency data transmission scheduling method for 5G edge computing according to claim 1, characterized in that, The data transmission task set is classified by latency sensitivity to generate a latency-critical task subset and a latency-tolerant task subset, including: Extract the task generation timestamp and preset latest completion deadline for each transmission task from the data transmission task set; Based on the task generation timestamp and the latest completion deadline, the maximum allowable transmission duration for each transmission task is calculated. The maximum allowed transmission time is compared with a preset latency sensitivity threshold, which is determined based on the network reference round-trip time and jitter tolerance. All transmission tasks whose maximum allowed transmission duration is less than the latency sensitivity threshold are classified as the latency-critical task subset, and the tasks in the latency-critical task subset have low tolerance for transmission interruption and delay. All transmission tasks whose maximum allowed transmission duration is greater than or equal to the latency-sensitive threshold are classified into the latency-tolerant task subset, and tasks in the latency-tolerant task subset are allowed to be completed within a larger time window.

3. The low-latency data transmission scheduling method for 5G edge computing according to claim 1, characterized in that, For the aforementioned delay-critical task subset, perform channel reservation processing based on predicted network state, including: Obtain a set of channel quality indicators for the current and historical moments, wherein the set of channel quality indicators includes reference signal received power, channel busy ratio, and bit error rate statistics; Acquire real-time network load data, which includes the number of active user devices and the average length of the data queue. The set of channel quality indicators and the real-time network load data are input into a pre-trained neural network model, and the neural network model outputs a predicted probability of channel availability for several future transmission time slots. Based on the predicted channel availability probability, channels with availability probabilities consistently higher than the reservation threshold are selected to calculate channel reservation schemes for tasks in the latency-critical task subset. The channel reservation scheme specifies the starting time slot, duration time slot, and channel number reserved for each task. The channel reservation scheme is distributed to relevant edge computing nodes and wireless access network devices to lock channel resources for a specified task in a specified time slot.

4. The low-latency data transmission scheduling method for 5G edge computing according to claim 1, characterized in that, For the aforementioned delay-tolerant task subset, perform batch transmission scheduling processing based on dynamic aggregation of available channels, including: Monitor the real-time status of all channels in the current network and record the predicted idle time window for each channel in the future; For the tasks in the latency-tolerant task subset, cluster them according to their source node, target node and data size, and aggregate multiple tasks with the same source node and target node and small data size into a batch transmission task. For each of the batch transmission tasks, one or more predicted idle time windows that are consecutive or adjacent in time are matched, such that the total length of the predicted idle time windows is sufficient to complete the transmission of the batch transmission task. Based on the matching results, a batch transmission scheduling plan is generated, which specifies the start transmission time, the channel used, and the transmission order of each original subtask within the task for each batch transmission task. When the predicted idle time window actually arrives, data transmission is activated according to the batch transmission scheduling plan, and channel quality is monitored during transmission. If the channel quality deteriorates to the switching threshold, a switch to a backup idle window or channel is triggered.

5. The low-latency data transmission scheduling method for 5G edge computing according to claim 4, characterized in that, The tightening adjustment of the reservation strategy includes increasing the channel resource margin reserved for future missions or selecting channels with higher predicted availability, including: Based on historical delay deviation values, the number of additional protection time slots required for future predictions is calculated, and the number of additional protection time slots is proportional to the average delay deviation value. When calculating the channel reservation scheme for new time-delay-critical tasks, the number of additional protection time slots is increased based on the minimum number of transmission time slots estimated according to the data volume and theoretical rate, and this number is used as the new total number of reserved time slots. Update the channel selection criteria: when selecting a target reserved channel from the candidate channel set, prioritize channels whose predicted availability probability exceeds a higher-level reservation threshold. When allocating reserved resources, a backup channel is reserved for the most critical tasks. The backup channel is activated when the quality of the primary channel deteriorates to the switching threshold.

6. The low-latency data transmission scheduling method for 5G edge computing according to claim 5, characterized in that, When the predicted idle time window actually arrives, data transmission is activated according to the batch transmission scheduling plan, and channel quality is monitored during transmission. If the channel quality deteriorates to the handover threshold, a handover to a backup idle window or channel is triggered, including: When the batch transmission task begins as scheduled, real-time channel quality monitoring is initiated, and the monitoring indicators include instantaneous bit error rate and signal-to-interference-to-noise ratio. When the instantaneous bit error rate is continuously exceeded by the bit error rate threshold, or the signal-to-interference-plus-noise ratio is continuously lower than the signal-to-interference-plus-noise ratio threshold, it is determined that the current channel quality has deteriorated to the point of reaching the handover threshold. Immediately suspend the current batch transmission task and query the backup idle time window or backup channel pre-allocated for the batch transmission task; If a spare idle time window is available and time-coordinated, the remaining untransmitted data will be scheduled to continue transmission within the spare idle time window. If no suitable alternative idle time window exists, but other idle channels are available, the remaining untransmitted data will be immediately switched to those other idle channels to continue transmission.

7. The low-latency data transmission scheduling method for 5G edge computing according to claim 3, characterized in that, The construction methods of the pre-trained neural network model include: Obtain a historical network status dataset, which includes channel quality metrics and real-time network load data collected over multiple historical transmission cycles, as well as actual channel status labels for subsequent historical transmission cycles. The actual channel status labels are used to identify whether the channel is available in a specific future time slot. The historical network state dataset is preprocessed, including filling in missing values, normalizing the data, and constructing model input samples according to a preset time series length. A neural network model is constructed, which adopts an architecture combining a long short-term memory network and an attention mechanism. The long short-term memory network is used to extract time-dependent features from the network state data, and the attention mechanism is used to weight and focus on key time period information in the time-dependent features. The preprocessed network state dataset is divided into a training set, a validation set, and a test set. The neural network model is trained using the training set. During the training process, the model hyperparameters are tuned using the validation set, and the model's prediction accuracy is evaluated using the test set. Once the model's prediction accuracy reaches a preset threshold, the model parameters are saved, completing the construction of the pre-trained neural network model.

8. The low-latency data transmission scheduling method for 5G edge computing according to claim 1, characterized in that, The method also includes an online learning step for the network state prediction model: Collect the actual execution results of channel reservation processing and batch transmission scheduling processing, including the actual channels used, actual transmission delay, and actual interference events encountered; The actual execution results are compared with the predicted network state on which the scheduling decision was based to generate a prediction error sample. The predicted error samples are used to incrementally train the pre-trained neural network model, and the internal weight parameters of the neural network model are updated. Before the start of the next transmission cycle, a new channel availability probability prediction is made using the updated neural network model.

Images