Method and apparatus for adaptive transmission in low-latency network for integrated services

Abstract

The application discloses a low-latency network transmission adaptive method and device for comprehensive services, and relates to the technical field of communication networks; the method comprises the following steps: receiving an acknowledgement message to extract ECN mark information and RTT measurement values, obtaining ECN intensity and normalized RTT gradient; accumulating the characteristics of multiple observation periods and inputting a CART decision tree to perform scene discrimination, obtaining a dominant scene category; selecting target parameters according to the scene category and performing smoothing to obtain an energy leakage factor and a back-off activation threshold; taking the energy leakage factor as a weight to fuse the ECN intensity and the RTT positive gradient to generate a congestion energy state; mapping the congestion energy state into a smooth growth gain and an adaptive back-off modulation factor, and selecting one of them to update a congestion window according to whether there is an ECN mark; taking the back-off activation threshold as a trigger judgment benchmark and returning to continue; the application can distinguish transient disturbance from persistent congestion, and realizes continuous and smooth adjustment of the congestion window.

Description

Technical Field

[0001] This invention relates to the field of communication network technology, and in particular to an adaptive method and apparatus for low-latency network transmission oriented towards integrated services. Background Technology

[0002] With the continuous development of heterogeneous power grid networking technology, heterogeneous power grid communication networks are gradually becoming an important infrastructure supporting new low-latency power services such as precise control of smart grids, remote operation and maintenance of power grids, and collaborative scheduling of power communication resources. These new power services not only require continuous and stable high data transmission rates, but also rely on the continuous assurance of end-to-end communication links in terms of transmission latency and operational stability. Power industry standards have clearly stipulated that the end-to-end transmission latency of core power services such as real-time grid dispatching and differential protection must be strictly controlled within 20 milliseconds to ensure rapid response to grid faults and safe and stable operation.

[0003] To reduce queuing latency in dedicated power grid networks and address jitter issues in power grid service transmission at the system level, the Internet Engineering Task Force (IETF) proposed the L4S architecture, which is adaptable to the power grid communication optimization requirements. The core of this architecture is end-to-end network collaborative optimization: on the network side, a dual-queue coupled active queue management (DualQ Coupled AQM) mechanism and an explicit congestion notification (ECN) marking mechanism are used to accurately mark and identify the link congestion status of the dedicated power grid; on the data sending end, a scalable congestion control algorithm (such as TCP Prague) is used to dynamically adjust the transmission rate of power service data based on the ECN congestion marking rate, adapting to the complex communication transmission scenarios of the power grid.

[0004] TCP Prague is the IETF's standard sender congestion control algorithm under the L4S architecture. Its design follows the core idea of ​​scaling backoff strength proportionally to the number of tags in data center transmission control protocols. Its main mechanism is as follows: the sender counts the proportion of packets marked by the ECN in each statistical period, and smooths it using an exponentially weighted moving average to obtain the smoothed tagging rate. Then follow the formula Perform a multiplicative rollback on the congestion window.

[0005] Existing technologies lack the ability to identify short-term wireless disturbances, making them prone to congestion misjudgments. In highly dynamic wireless environments such as 5G / 6G, link states are susceptible to frequent fluctuations due to factors such as channel fading, scheduling changes, and transient wireless disturbances. While the Hybrid Automatic Repeat reQuest (HARQ) mechanism improves transmission reliability, it can introduce brief pauses and random latency jitter. TCP Prague misjudges these short-term wireless disturbances as persistent congestion, triggering unnecessary window backoffs, leading to decreased throughput and performance oscillations. Tests conducted by Brunello et al. in a 5G simulation environment show that TCP Prague is overly sensitive to wireless fluctuations.

[0006] Existing technologies lack sufficient perception of network state change trends and are deficient in scenario adaptation capabilities. TCP Prague's window adjustment mainly relies on the ECN tagging rate, which reflects more the strength of congestion signals within the current period and provides relatively limited information on whether the network state is continuously deteriorating or only experiencing short-term fluctuations. When the dominant network mode (such as noise-dominated, burst-dominated, or congestion-dominated) changes over time, fixed-parameter controllers struggle to maintain optimal performance across different scenarios.

[0007] Existing congestion control algorithms with fixed parameters struggle to balance stability and response speed across different scenarios. Overly conservative parameter settings lead to slow responses to real congestion, while overly sensitive parameters can easily misinterpret transient noise or sudden service bursts as congestion. Current methods lack adaptive mechanisms to dynamically adjust control parameters based on network conditions. Summary of the Invention

[0008] This invention provides an adaptive method and apparatus for low-latency network transmission for integrated services, which solves the problem that fixed-parameter congestion control algorithms in the prior art cannot balance stability and response speed under different network scenarios, and realizes scenario adaptive identification of network status and dynamic adjustment of control parameters.

[0009] In a first aspect, the present invention provides an adaptive method for low-latency network transmission oriented towards integrated services, comprising: S101, Receive an acknowledgment message, extract ECN tag information and RTT measurement value from the acknowledgment message, and obtain the ECN strength and normalized RTT gradient of the current period; S102, accumulate RTT measurement values ​​and ECN label information from multiple observation periods and extract observable feature vectors, input the observable feature vectors into the CART decision tree for scene discrimination, and obtain the dominant scene category of the current network. S103, select the corresponding target parameters from the preset parameter set according to the dominant scenario category, and perform exponential weighted moving average smoothing on the target parameters to obtain the currently effective energy leakage factor and rollback activation threshold; S104, using the energy leakage factor as the historical state retention weight, the current period's ECN intensity and the normalized RTT positive gradient are weighted and fused to generate the congestion energy state. S105, the congestion energy state is mapped to a smooth growth gain and an adaptive backoff modulation factor respectively. The smooth growth gain or the adaptive backoff modulation factor is selected according to whether there is an ECN mark in the current period to update the current congestion window, and the updated congestion window is obtained. S106, using the backoff activation threshold as the triggering criterion for the adaptive backoff modulation factor, and controlling the transmission of subsequent data packets with the updated congestion window, then returning to S102.

[0010] In conjunction with the first aspect, in one possible implementation, the step of extracting ECN tag information and RTT measurement values ​​from the acknowledgment message to obtain the ECN strength and normalized RTT gradient for the current period includes: Set a sliding observation window at the transmitting end The proportion of packets marked with ECN within the statistical window is used to obtain the ECN strength; Smoothing the instantaneous RTT to obtain ,set up Calculate the normalized RTT gradient to obtain the minimum RTT since the connection was established.

[0011] In conjunction with the first aspect, in one possible implementation, the accumulation of RTT measurements and ECN tagging information across multiple observation periods and the extraction of observable feature vectors includes: Within each sliding observation window, the proportion of packets marked as CE to the total number of packets is counted to obtain the ECN marking density; Within the sliding observation window, each data packet is represented as a binary sequence according to the arrival order of the acknowledgment messages, indicating whether it carries the CE tag. The longest consecutive sequence of 1s in the binary sequence is counted to obtain the maximum consecutive length. Within the sliding observation window, the standard deviation of all RTT measurements is calculated, and the standard deviation is divided by the smaller of the minimum RTT since the connection was established and the mean RTT in the current window to obtain the normalized RTT jitter. Perform temporal differencing on the smoothed RTT and extract the positive portion of the differencing result as the positive gradient energy; The ECN label density, the maximum continuous length, the normalized RTT jitter, and the positive gradient energy are used as the observable feature vector.

[0012] In conjunction with the first aspect, in one possible implementation, the step of inputting the observable feature vector into a CART decision tree for scene discrimination to obtain the dominant scene category of the current network includes: The CART decision tree is a classification model pre-trained with Gini impurity as the splitting criterion. The leaf node output categories of the classification model include noise-dominated mode, burst-dominated mode, and congestion-dominated mode. After the observable feature vector is input into the CART decision tree, it is traversed layer by layer down along the branches of the CART decision tree until a leaf node is reached. The output category corresponding to the leaf node is taken as the dominant scene category of the current network. The noise-dominated mode characterizes short-term non-congestion fluctuations caused by physical layer wireless channel disturbances, the burst-dominated mode characterizes short-term queuing processes caused by application layer service bursts, and the congestion-dominated mode characterizes a persistent congestion state caused by link overload.

[0013] In conjunction with the first aspect, in one possible implementation, the step of selecting corresponding target parameters from a preset parameter set based on the dominant scenario category, and performing exponentially weighted moving average smoothing on the target parameters to obtain the currently effective energy leakage factor and fallback activation threshold includes: The preset parameter set includes preset values ​​for energy leakage factors and preset values ​​for backoff activation thresholds corresponding to noise-dominant mode, burst-dominant mode and congestion-dominant mode, respectively. When the noise-dominated mode is identified, the maximum preset value of the energy leakage factor and the maximum preset value of the backoff activation threshold are selected as target parameters. When the sudden dominant mode is identified, the preset values ​​of the energy leakage factor and the back-off activation threshold are selected as the target parameters. When the congestion-dominated mode is identified, the minimum preset value of the energy leakage factor and the minimum preset value of the backoff activation threshold are selected as target parameters. The energy leakage factor that is actually effective in the current period is obtained by weighting the actual effective parameters of the previous period and the target parameters of the current period with a preset smoothing factor. The actual effective rollback activation threshold of the previous period and the target rollback activation threshold of the current period are weighted and averaged using the preset smoothing factor to obtain the actual effective rollback activation threshold of the current period.

[0014] In conjunction with the first aspect, in one possible implementation, the step of using the energy leakage factor as a historical state retention weight to weight and fuse the current period's ECN intensity with the normalized RTT positive gradient to generate a congestion energy state includes: Multiply the congestion energy state of the previous cycle by the energy leakage factor to obtain the historical energy retention value; The current ECN energy injection value is obtained by multiplying the current cycle's ECN strength by one and subtracting the energy leakage factor. Multiply the normalized RTT positive gradient of the current cycle by the preset gradient weight coefficient to obtain the gradient energy injection value; The historical energy retention value, the current ECN energy injection value, and the gradient energy injection value are added together, and the sum is used as the congestion energy state for the current period.

[0015] In conjunction with the first aspect, in one possible implementation, mapping the congestion energy state to a smooth growth gain and an adaptive backoff modulation factor, and updating the current congestion window by selecting either the smooth growth gain or the adaptive backoff modulation factor based on whether an ECN marker exists in the current period, to obtain an updated congestion window, includes: The congestion energy state is input into the first Sigmoid function. The first Sigmoid function outputs a gain that approaches the maximum value when the congestion energy state is lower than the first preset threshold and outputs a gain that approaches zero when the congestion energy state is higher than the first preset threshold. The output of the first Sigmoid function is used as the smooth growth gain. The congestion energy state is input into the second Sigmoid function. The second Sigmoid function outputs a modulation value approaching zero when the congestion energy state is lower than the second preset threshold and outputs a modulation value approaching one when the congestion energy state is higher than the second preset threshold. The output of the second Sigmoid function is used as the adaptive backoff modulation factor. If an ECN tag exists in the current period, subtract half of the product of the current congestion window, the smoothing tag rate, and the adaptive backoff modulation factor from the current congestion window, and use the result of the subtraction as the updated congestion window. If there is no ECN marker in the current period, add the smooth growth gain to the current congestion window and divide it by the quotient of the current congestion window, and use the sum as the updated congestion window.

[0016] In conjunction with the first aspect, in one possible implementation, using the back-off activation threshold as the triggering criterion for the adaptive back-off modulation factor includes: When the congestion energy state is less than the back-off activation threshold, the adaptive back-off modulation factor takes a value less than 0.5; When the congestion energy state is greater than the back-off activation threshold, the adaptive back-off modulation factor takes a value greater than 0.5.

[0017] Secondly, the present invention provides a low-latency network transmission adaptive device for integrated services, comprising: The multidimensional signal observation module is used to receive acknowledgment messages, extract ECN tag information and RTT measurement values ​​from the acknowledgment messages, and output the ECN intensity and normalized RTT gradient of the current period. The scene feature extraction and recognition module is used to accumulate RTT measurement values ​​and ECN label information from multiple observation periods and extract observable feature vectors. The observable feature vectors are then input into the CART decision tree for scene discrimination, and the dominant scene category of the current network is output. The parameter scheduling and smoothing update module is used to select the corresponding target parameters from the preset parameter set according to the dominant scenario category, perform exponential weighted moving average smoothing on the target parameters, and output the currently effective energy leakage factor and rollback activation threshold. The congestion energy integration module is used to receive the ECN intensity and normalized RTT gradient output by the multidimensional signal observation module, as well as the energy leakage factor output by the parameter scheduling and smoothing update module. The energy leakage factor is used as the historical state retention weight to perform weighted fusion of the current period's ECN intensity and normalized RTT positive gradient, and output the congestion energy state. The dual-mode control module is used to receive the congestion energy state output by the congestion energy integration module and the back-off activation threshold output by the parameter scheduling and smooth update module, respectively map the congestion energy state to the smooth growth gain and the adaptive back-off modulation factor, and select the smooth growth gain or the adaptive back-off modulation factor to update the current congestion window according to whether there is an ECN flag in the current period, and output the updated congestion window. The closed-loop feedback module is used to apply the updated congestion window output by the dual-mode control module to the transmission of subsequent data packets and to trigger the scene feature extraction and recognition module to enter the next observation cycle.

[0018] One or more technical solutions provided in this invention have at least the following technical effects or advantages: This invention receives acknowledgment messages and extracts ECN tag information and RTT measurement values ​​from them to obtain the ECN strength and normalized RTT gradient for the current period. The beneficial effect of this step is that by directly extracting ECN tags and RTT measurement values ​​in parallel from acknowledgment messages, it achieves synchronous perception of network congestion levels and trends. The ECN strength can instantly reflect the cumulative degree of explicit congestion notifications within the current period, while the normalized RTT gradient effectively characterizes the transient change direction of round-trip time. The combination of these two provides high-quality raw input with both instantaneous state and changing trends for subsequent decision-making, while avoiding the bandwidth overhead of additional probe messages. It accumulates RTT measurement values ​​and ECN tag information over multiple observation periods. The process involves extracting observable feature vectors and inputting them into a CART decision tree for scene discrimination to determine the dominant scene category of the current network. The beneficial effect of this step is that by accumulating observable feature vectors across cycles, it fully exploits the time-series correlation information of the network state. Simultaneously, leveraging the advantages of CART decision trees—strong non-linear classification ability, low computational complexity, and easily interpretable results—it efficiently maps complex, multi-dimensional input features into a finite and clear set of scene categories. This accurately distinguishes between shallow-threshold ECN-labeled scenes, deep-queue packet loss scenes, burst interference scenes in wireless links, and network idle scenes, providing a reliable basis for subsequent adaptive parameter selection and significantly improving the algorithm's generalization ability in different network environments. Robustness; Based on the dominant scenario category, corresponding target parameters are selected from the preset parameter set, and the target parameters are smoothed by exponentially weighted moving average to obtain the currently effective energy leakage factor and backoff activation threshold; The beneficial effect of this step is that, based on the scenario discrimination result, the control parameters are dynamically matched on demand, so that the energy leakage factor and backoff activation threshold can be adaptively adjusted with changes in the macro-environment of the network. At the same time, the introduction of exponentially weighted moving average smoothing effectively suppresses the drastic jitter of parameters at the scenario switching boundary, avoids congestion window oscillation caused by parameter mutation, and ensures the continuity and stability of control parameters, thus achieving a good balance between rapid response to scenario changes and maintaining control stability; using the energy leakage factor as The historical state retention weighting method involves weighted fusion of the current period's ECN strength and the normalized RTT positive gradient to generate a congestion energy state. The beneficial effect of this step is that it assigns a controllable decay weight to historical congestion information through an energy leakage factor, so that the generated congestion energy state can not only retain the long-term memory of historical cumulative congestion effects, but also dynamically adjust the retention degree of historical information according to the current network scenario. At the same time, it weighted fusion of the current congestion severity represented by the ECN strength and the congestion deterioration trend represented by the normalized RTT positive gradient to construct a single and dimensionally unified congestion energy state quantity, which effectively integrates the magnitude information and change rate information of congestion, providing a more comprehensive and accurate state benchmark for subsequent congestion window control than a single indicator.The congestion energy state is mapped to a smooth growth gain and an adaptive backoff modulation factor. Depending on the presence of an ECN marker in the current period, either the smooth growth gain or the adaptive backoff modulation factor is selected to update the current congestion window, resulting in the updated congestion window. The advantage of this step is that it generates two independent sets of window adjustment parameters based on the congestion energy state. The smooth growth gain is used during the window expansion phase without an ECN marker to ensure bandwidth utilization efficiency, while the adaptive backoff modulation factor is used during the window contraction phase with an ECN marker to achieve rapid congestion mitigation. Activation is selectively based on the presence or absence of an ECN marker in the current period, achieving decoupling and independent optimization of the growth and backoff modes. This avoids the performance loss caused by compromising between the two modes with a single adjustment factor. Furthermore, the mapping relationship between the congestion energy state and the adjustment factor ensures that the window update amplitude can continuously and smoothly change with the congestion level, avoiding the limitations of traditional additive methods. The fixed-step-size mutation problem in the multiplicative-subtractive algorithm is addressed. Using a backoff activation threshold as the trigger for the adaptive backoff modulation factor, and controlling the transmission of subsequent data packets with an updated congestion window, this step effectively avoids unnecessary throughput loss due to excessive backoff when congestion energy is low. It ensures that backoff operations are only activated when congestion energy truly accumulates to a level requiring intervention. Simultaneously, the updated congestion window precisely controls the transmission rate and quantity of subsequent data packets, directly applying algorithmic decisions to actual data flow transmission behavior, forming a complete closed-loop control link of perception, decision-making, execution, and feedback. Furthermore, by returning to S102, the observation window continuously slides forward, enabling the entire congestion control process to iteratively optimize in each cycle, maintaining high-frequency tracking and timely response to the dynamic network environment. Attached Figure Description

[0019] Figure 1 A flowchart illustrating the steps of an adaptive low-latency network transmission method for integrated services provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the congestion energy integrator signal processing provided in an embodiment of the present invention; Figure 3 A schematic diagram of a dual-mode Sigmoid control law provided in an embodiment of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0021] In a first aspect, the present invention provides an adaptive method for low-latency network transmission oriented towards integrated services, see [link to relevant documentation]. Figure 1 ,include: S101, Receive the acknowledgment message, extract the ECN tag information and RTT measurement value from the acknowledgment message, and obtain the ECN strength and normalized RTT gradient of the current period; Specifically, in step S101, ECN tag information and RTT measurement value are extracted from the acknowledgment message to obtain the ECN strength and normalized RTT gradient of the current period, including the following steps S1011 to S1012.

[0022] S1011, Set a sliding observation window at the transmitting end. The proportion of packets marked with ECN within the statistical window is used to obtain the ECN strength; S1012, smoothing the instantaneous RTT to obtain ,set up Calculate the normalized RTT gradient to obtain the minimum RTT since the connection was established.

[0023] For example, a sliding observation window is set at the sending end. The intensity of ECN marked with ECN within the statistical window is represented as follows: ; in, Indicates the observation window The total number of data units confirmed within the scope, As an indicator function, when the data packet carries The value is 1 when it is marked, and 0 otherwise.

[0024] Smoothing the instantaneous RTT to obtain ,set up The normalized rate of change is defined as the minimum RTT since the connection was established: ; when This indicates that RTT is continuously rising and the queue is accumulating; when This indicates that the queue length is stabilizing or diminishing.

[0025] S102: Accumulate RTT measurement values ​​and ECN label information from multiple observation periods and extract observable feature vectors. Input the observable feature vectors into the CART decision tree for scene discrimination to obtain the dominant scene category of the current network. Specifically, in step S102, the RTT measurement values ​​and ECN label information of multiple observation periods are accumulated and the observable feature vector is extracted, including the following steps S1021 to S1025.

[0026] S1021, within each sliding observation window, count the proportion of packets marked as CE to the total number of packets to obtain the ECN marking density; S1022, within the sliding observation window, represent each data packet as a binary sequence according to the arrival order of the acknowledgment messages, indicating whether each data packet carries the CE tag. Calculate the longest consecutive 1 in the binary sequence to obtain the maximum consecutive length. S1023, within the sliding observation window, calculate the standard deviation of all RTT measurements, and divide the standard deviation by the smaller of the minimum RTT since the connection was established and the mean RTT in the current window to obtain the normalized RTT jitter. S1024, perform time-series difference on smoothed RTT, and extract the positive part of the difference result as positive gradient energy; S1025 uses ECN label density, maximum continuous length, normalized RTT jitter, and positive gradient energy as observable feature vectors.

[0027] Specifically, in step S102, the observable feature vector is input into the CART decision tree for scene discrimination to obtain the dominant scene category of the current network, including: (1) The CART decision tree is a classification model that is pre-trained with Gini impurity as the splitting criterion. The leaf node output categories of the classification model include noise-dominated mode, burst-dominated mode and congestion-dominated mode. (2) After inputting the observable feature vector into the CART decision tree, the CART decision tree is traversed layer by layer down the branches until the leaf node is reached. The output category corresponding to the leaf node is taken as the dominant scene category of the current network. (3) Noise-dominated mode represents short-term non-congestion fluctuations caused by physical layer wireless channel disturbances, burst-dominated mode represents short-term queuing processes caused by application layer service bursts, and congestion-dominated mode represents continuous congestion caused by link overload.

[0028] For example, observable feature vector extraction includes extracting the following feature vectors over a sliding window: ECN marker density Represented as: ; Maximum continuous length It is the longest consecutive sequence of 1s in the CE marker indicator sequence arranged in the ACK sequence within the window, and the maximum consecutive length. Represented as: ; Normalized RTT jitter Represented as: ; Positive gradient energy Represented as: ; in, ; The observable feature vector is: .

[0029] Using CART decision trees as the scene recognition model, the network state is divided into three dominant modes: Noise-dominated mode is characterized by short-term disturbances caused by physical layer factors such as wireless channel fading and HARQ retransmission. It has large RTT jitter, discrete and sporadic ECN markers, and no sustained positive gradient.

[0030] Burst mode is characterized by short-term queuing caused by application layer bursts such as XR service I-frame pulses, local concentration of ECN markers, and rapid decay of the RTT positive gradient.

[0031] Congestion-dominated mode is characterized by persistent congestion caused by real link overload, with a continuous positive gradient in RTT and continuous high density of ECN markers.

[0032] The CART decision tree is trained using Gini impurity as the node splitting criterion, and the maximum tree depth is limited to 5 to ensure low latency for online inference. The inference time complexity is O(n log n). , With a tree depth, scene recognition can be completed with only a finite number of threshold comparisons.

[0033] S103: Select the corresponding target parameters from the preset parameter set according to the dominant scenario category, and perform exponential weighted moving average smoothing on the target parameters to obtain the currently effective energy leakage factor and rollback activation threshold. Specifically, in step S103, the corresponding target parameter is selected from the preset parameter set according to the dominant scenario category, and the target parameter is smoothed by exponential weighted moving average to obtain the currently effective energy leakage factor and the rollback activation threshold, including the following steps S1031 to S1036.

[0034] S1031, The preset parameter set includes preset values ​​of energy leakage factor and preset values ​​of back-off activation threshold for noise-dominated mode, burst-dominated mode and congestion-dominated mode respectively. S1032, When the noise-dominated mode is identified, the maximum preset value of the energy leakage factor and the maximum preset value of the back-off activation threshold are selected as target parameters; S1033, when the sudden dominant mode is identified, the preset value of the energy leakage factor and the preset value of the back-off activation threshold are selected as the target parameters. S1034, When the congestion-dominated mode is identified, the minimum preset value of the energy leakage factor and the minimum preset value of the backoff activation threshold are selected as target parameters; S1035, use a preset smoothing factor to perform a weighted average of the actual effective parameters of the previous cycle and the target parameters of the current cycle to obtain the energy leakage factor that is actually effective in the current cycle. S1036, using a preset smoothing factor, the actual effective rollback activation threshold of the previous period and the target rollback activation threshold of the current period are weighted and averaged to obtain the actual effective rollback activation threshold of the current period.

[0035] For example, based on the identified dominant scene category The target parameters are obtained by looking up the table.

[0036]

[0037] To achieve a smooth transition of parameters, an exponentially weighted moving average is used for updates. The actual effective parameters for the current period are represented as follows: The energy leakage factor that is actually effective in the current cycle is expressed as: ; The actual rollback activation threshold for the current period is expressed as follows: ; in, Smoothing factor (preferred in this invention) This allows the control parameters to smoothly transition to the new target value within approximately 15 ACK cycles, avoiding window oscillations caused by parameter mutations.

[0038] S104 uses the energy leakage factor as the historical state retention weight, and weights and fuses the current period's ECN intensity with the normalized RTT positive gradient to generate the congestion energy state. Specifically, in step S104, the energy leakage factor is used as the historical state retention weight, and the current period's ECN intensity is weighted and fused with the normalized RTT positive gradient to generate the congestion energy state, including the following steps S1041 to S1044.

[0039] S1041, multiply the congestion energy state of the previous cycle by the energy leakage factor to obtain the historical energy retention value; S1042, Multiply the current period's ECN strength by the difference between one and the energy leakage factor to obtain the current ECN energy injection value; S1043, multiply the normalized RTT positive gradient of the current period by the preset gradient weight coefficient to obtain the gradient energy injection value; S1044 adds the historical energy retention value, the current ECN energy injection value, and the gradient energy injection value together, and uses the sum as the congestion energy state for the current cycle.

[0040] For example, see Figure 2 The ECN intensity and RTT gradient are fused into the congestion energy state for the current cycle. The core update equation is: ; in, Energy leakage factor ( This determines the degree to which historical congestion information is retained; The gradient weight coefficient adjusts the strength of the influence of the RTT trend term on the energy state.

[0041] This integrator includes three functions: Integral term The memory units that constitute the system require a continuous energy injection to change the system state, thus preventing transient signal jumps from directly triggering window rollback. Proportional Term : Directly coupled with the current ECN observation intensity, energy accumulates steadily as congestion persists; Differential term Introducing the RTT change trend allows for early energy injection to compensate for phase lag in the integral phase, as congestion is formed in the early stages.

[0042] S105, the congestion energy state is mapped to a smooth growth gain and an adaptive backoff modulation factor respectively. Based on whether there is an ECN marker in the current period, the smooth growth gain or the adaptive backoff modulation factor is selected to update the current congestion window, and the updated congestion window is obtained. Specifically, in step S105, the congestion energy state is mapped to a smooth growth gain and an adaptive backoff modulation factor, respectively. The smooth growth gain or the adaptive backoff modulation factor is selected to update the current congestion window according to whether there is an ECN mark in the current period, and the updated congestion window is obtained, including the following steps S1051 to S1054.

[0043] S1051, The congestion energy state is input into the first Sigmoid function. The first Sigmoid function outputs a gain that approaches the maximum value when the congestion energy state is lower than the first preset threshold and outputs a gain that approaches zero when the congestion energy state is higher than the first preset threshold. The output of the first Sigmoid function is used as the smooth growth gain. S1052, The congestion energy state is input into the second Sigmoid function. The second Sigmoid function outputs a modulation value approaching zero when the congestion energy state is lower than the second preset threshold and outputs a modulation value approaching one when the congestion energy state is higher than the second preset threshold. The output of the second Sigmoid function is used as the adaptive backoff modulation factor. S1053, If an ECN mark exists in the current period, subtract half of the product of the current congestion window, the smoothing mark rate, and the adaptive backoff modulation factor from the current congestion window, and use the result of the subtraction as the updated congestion window. S1054, if there is no ECN marker in the current period, add the smooth growth gain to the current congestion window and divide it by the quotient of the current congestion window, and use the sum as the updated congestion window.

[0044] For example, see Figure 3 In the dual-mode Sigmoid controller, during the window growth phase, a nonlinear gain scheduling based on congestion energy is employed to smoothly increase the gain. Specifically, it is expressed as follows: ; in, For maximum growth gain, The growth decay threshold, Control the steepness of the curve. Low energy region. Allows for faster growth to utilize idle bandwidth; high-energy areas The window growth slows down smoothly and eventually stops.

[0045] Adaptive back-off modulation factor Represented as: ; in, To roll back the activation threshold, Control the speed of rollback switching. It exhibits S-shaped curve characteristics: Noise shielding area ( ): Even with ECN echo, the window back strength is limited to a small range; Fast transition zone ( The regression factor is rising rapidly, prompting the initiation of gentle intervention. Saturation region ( ): It fully restores the deceleration intensity of the standard Prague.

[0046] When a congestion flag is detected, i.e. : ; When no congestion marker is detected, i.e. : .

[0047] S106, using the back-off activation threshold as the triggering criterion for the adaptive back-off modulation factor, and controlling the transmission of subsequent data packets with the updated congestion window, then returning to S102.

[0048] Specifically, in step S106, the back-off activation threshold is used as the triggering criterion for the adaptive back-off modulation factor, including the following steps S1061 to S1062.

[0049] S1061, When the congestion energy state is less than the backoff activation threshold, the adaptive backoff modulation factor is less than 0.5; S1062, when the congestion energy state is greater than the backoff activation threshold, the adaptive backoff modulation factor takes a value greater than 0.5.

[0050] For example, when the congestion energy state Less than the fallback activation threshold At that time, the adaptive back-off modulation factor A value less than 0.5 indicates that the current congestion energy has not reached the level that would trigger a strong backoff, and the window backoff strength is limited to a small range. When the congestion energy state Greater than the fallback activation threshold At that time, the adaptive back-off modulation factor A value greater than 0.5 indicates that the current congestion energy has exceeded the trigger threshold, and the window back-off strength increases with the increase of congestion energy; After completing the congestion window update for the current period, the transmission rate of subsequent data packets is controlled by the updated congestion window, and the process returns to step S102. The observable feature vector is re-extracted using the acknowledgment messages collected in the next observation period, forming a closed-loop control process of continuous scene recognition, parameter adjustment, congestion energy integration, dual-mode mapping, and window update.

[0051] Secondly, the present invention provides a low-latency network transmission adaptive device for integrated services, comprising: The multidimensional signal observation module is used to receive acknowledgment messages, extract ECN tag information and RTT measurement values ​​from the acknowledgment messages, and output the ECN intensity and normalized RTT gradient of the current period. The scene feature extraction and recognition module is used to accumulate RTT measurement values ​​and ECN label information from multiple observation periods and extract observable feature vectors. The observable feature vectors are then input into the CART decision tree for scene discrimination, and the dominant scene category of the current network is output. The parameter scheduling and smoothing update module is used to select the corresponding target parameters from the preset parameter set according to the dominant scenario category, perform exponential weighted moving average smoothing on the target parameters, and output the currently effective energy leakage factor and rollback activation threshold. The congestion energy integration module receives the ECN strength and normalized RTT gradient output by the multidimensional signal observation module, as well as the energy leakage factor output by the parameter scheduling and smoothing update module. Using the energy leakage factor as the historical state retention weight, the current period's ECN strength and normalized RTT positive gradient are weighted and fused to output the congestion energy state. The dual-mode control module receives the congestion energy state output by the congestion energy integration module and the back-off activation threshold output by the parameter scheduling and smooth update module. It maps the congestion energy state to a smooth growth gain and an adaptive back-off modulation factor, respectively. Based on whether there is an ECN marker in the current period, it selects the smooth growth gain or the adaptive back-off modulation factor to update the current congestion window and outputs the updated congestion window. The closed-loop feedback module is used to apply the updated congestion window output by the dual-mode control module to the subsequent data packet transmission and trigger the scene feature extraction and recognition module to enter the next observation cycle.

[0052] For example, in actual operation, the multi-dimensional signal observation module continuously receives acknowledgment messages from the network layer, parses the ECN flag and RTT timestamp from each acknowledgment message, and outputs the calculated ECN strength and normalized RTT gradient to the congestion energy integration module at the end of each observation window. The scene feature extraction and recognition module extracts four observable features at the end of each observation window, inputs them into a CART decision tree to obtain the dominant scene category, and outputs them to the parameter scheduling and smoothing update module. The parameter scheduling and smoothing update module looks up the target parameters according to the scene category, smooths them using an exponentially weighted moving average, and outputs the actually effective energy leakage factor to the congestion energy integration module, and outputs the actually effective backoff excitation... The live threshold is output to the dual-mode control module; the congestion energy integration module generates a congestion energy state by weighted fusion of ECN strength and normalized RTT positive gradient based on the energy leakage factor, and outputs it to the dual-mode control module; the dual-mode control module maps the congestion energy state to a smooth growth gain and an adaptive backoff modulation factor, respectively, and selects the corresponding control quantity to update the congestion window according to whether there is an ECN marker in the current period, and outputs the updated congestion window to the protocol stack of the transmitting end; the closed-loop feedback module applies the updated congestion window to the transmission of subsequent data packets, and triggers the scene feature extraction and recognition module to enter the next observation window, realizing a complete closed loop from signal observation, scene recognition, parameter scheduling, energy integration, dual-mode control to window update.

[0053] The various embodiments described in this specification are presented in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on its differences from other embodiments. All or part of this invention can be used in numerous general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, mobile communication terminals, multiprocessor systems, microprocessor-based systems, programmable electronic devices, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices, etc.

[0054] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the present invention.

Claims

1. A low-latency network transmission adaptive method for integrated services, characterized in that, include: S101, Receive an acknowledgment message, extract ECN tag information and RTT measurement value from the acknowledgment message, and obtain the ECN strength and normalized RTT gradient of the current period; S102, accumulate RTT measurement values ​​and ECN label information from multiple observation periods and extract observable feature vectors, input the observable feature vectors into the CART decision tree for scene discrimination, and obtain the dominant scene category of the current network. S103, select the corresponding target parameters from the preset parameter set according to the dominant scenario category, and perform exponential weighted moving average smoothing on the target parameters to obtain the currently effective energy leakage factor and rollback activation threshold; S104, using the energy leakage factor as the historical state retention weight, the current period's ECN intensity and the normalized RTT positive gradient are weighted and fused to generate the congestion energy state. S105, the congestion energy state is mapped to a smooth growth gain and an adaptive backoff modulation factor respectively. The smooth growth gain or the adaptive backoff modulation factor is selected according to whether there is an ECN mark in the current period to update the current congestion window, and the updated congestion window is obtained. S106, using the backoff activation threshold as the triggering criterion for the adaptive backoff modulation factor, and controlling the transmission of subsequent data packets with the updated congestion window, then returning to S102.

2. The adaptive low-latency network transmission method for integrated services according to claim 1, characterized in that, The step of extracting ECN tag information and RTT measurement values ​​from the acknowledgment message to obtain the ECN strength and normalized RTT gradient for the current period includes: Set a sliding observation window at the transmitting end The proportion of packets marked with ECN within the statistical window is used to obtain the ECN strength; Smoothing the instantaneous RTT to obtain ,set up Calculate the normalized RTT gradient to obtain the minimum RTT since the connection was established.

3. The adaptive low-latency network transmission method for integrated services according to claim 1, characterized in that, The process of accumulating RTT measurements and ECN labeling information from multiple observation periods and extracting observable feature vectors includes: Within each sliding observation window, the proportion of packets marked as CE out of the total number of packets is counted to obtain the ECN marking density; Within the sliding observation window, each data packet is represented as a binary sequence according to the arrival order of the acknowledgment messages, indicating whether it carries the CE tag. The longest consecutive sequence of 1s in the binary sequence is counted to obtain the maximum consecutive length. Within the sliding observation window, the standard deviation of all RTT measurements is calculated, and the standard deviation is divided by the smaller of the minimum RTT since the connection was established and the mean RTT in the current window to obtain the normalized RTT jitter. Perform temporal differencing on the smoothed RTT and extract the positive portion of the differencing result as the positive gradient energy; The ECN label density, the maximum continuous length, the normalized RTT jitter, and the positive gradient energy are used as the observable feature vector.

4. The adaptive low-latency network transmission method for integrated services according to claim 1, characterized in that, The step of inputting the observable feature vector into the CART decision tree for scene discrimination to obtain the dominant scene category of the current network includes: The CART decision tree is a classification model pre-trained with Gini impurity as the splitting criterion. The leaf node output categories of the classification model include noise-dominated mode, burst-dominated mode, and congestion-dominated mode. After the observable feature vector is input into the CART decision tree, it is traversed layer by layer down along the branches of the CART decision tree until a leaf node is reached. The output category corresponding to the leaf node is taken as the dominant scene category of the current network. The noise-dominated mode characterizes short-term non-congestion fluctuations caused by physical layer wireless channel disturbances, the burst-dominated mode characterizes short-term queuing processes caused by application layer service bursts, and the congestion-dominated mode characterizes a persistent congestion state caused by link overload.

5. The adaptive low-latency network transmission method for integrated services according to claim 1, characterized in that, The step of selecting corresponding target parameters from a preset parameter set according to the dominant scenario category, and performing exponentially weighted moving average smoothing on the target parameters to obtain the currently effective energy leakage factor and rollback activation threshold includes: The preset parameter set includes preset values ​​for energy leakage factors and preset values ​​for backoff activation thresholds corresponding to noise-dominant mode, burst-dominant mode and congestion-dominant mode, respectively. When the noise-dominated mode is identified, the maximum preset value of the energy leakage factor and the maximum preset value of the backoff activation threshold are selected as target parameters. When the sudden dominant mode is identified, the preset values ​​of the energy leakage factor and the back-off activation threshold are selected as the target parameters. When the congestion-dominated mode is identified, the minimum preset value of the energy leakage factor and the minimum preset value of the backoff activation threshold are selected as target parameters. The energy leakage factor that is actually effective in the current period is obtained by weighting the actual effective parameters of the previous period and the target parameters of the current period with a preset smoothing factor. The actual effective rollback activation threshold of the previous period and the target rollback activation threshold of the current period are weighted and averaged using the preset smoothing factor to obtain the actual effective rollback activation threshold of the current period.

6. The adaptive low-latency network transmission method for integrated services according to claim 1, characterized in that, The step of using the energy leakage factor as the historical state retention weight to weight and fuse the current period's ECN strength with the normalized RTT positive gradient to generate a congestion energy state includes: Multiply the congestion energy state of the previous cycle by the energy leakage factor to obtain the historical energy retention value; The current ECN energy injection value is obtained by multiplying the current cycle's ECN strength by one and subtracting the energy leakage factor. Multiply the normalized RTT positive gradient of the current cycle by the preset gradient weight coefficient to obtain the gradient energy injection value; The historical energy retention value, the current ECN energy injection value, and the gradient energy injection value are added together, and the sum is used as the congestion energy state for the current period.

7. The adaptive low-latency network transmission method for integrated services according to claim 1, characterized in that, The process involves mapping the congestion energy state to a smooth growth gain and an adaptive backoff modulation factor, respectively. Based on the presence of an ECN marker in the current period, the smooth growth gain or the adaptive backoff modulation factor is selected to update the current congestion window, resulting in an updated congestion window. This includes: The congestion energy state is input into the first Sigmoid function. The first Sigmoid function outputs a gain that approaches the maximum value when the congestion energy state is lower than the first preset threshold and outputs a gain that approaches zero when the congestion energy state is higher than the first preset threshold. The output of the first Sigmoid function is used as the smooth growth gain. The congestion energy state is input into the second Sigmoid function. The second Sigmoid function outputs a modulation value approaching zero when the congestion energy state is lower than the second preset threshold and outputs a modulation value approaching one when the congestion energy state is higher than the second preset threshold. The output of the second Sigmoid function is used as the adaptive backoff modulation factor. If an ECN tag exists in the current period, subtract half of the product of the current congestion window, the smoothing tag rate, and the adaptive backoff modulation factor from the current congestion window, and use the result of the subtraction as the updated congestion window. If there is no ECN marker in the current period, add the smooth growth gain to the current congestion window and divide it by the quotient of the current congestion window, and use the sum as the updated congestion window.

8. The adaptive low-latency network transmission method for integrated services according to claim 1, characterized in that, The step of using the back-off activation threshold as the triggering criterion for the adaptive back-off modulation factor includes: When the congestion energy state is less than the back-off activation threshold, the adaptive back-off modulation factor takes a value less than 0.5; When the congestion energy state is greater than the back-off activation threshold, the adaptive back-off modulation factor takes a value greater than 0.

5.

9. A low-latency network transmission adaptive device for integrated services, characterized in that, include: The multidimensional signal observation module is used to receive acknowledgment messages, extract ECN tag information and RTT measurement values ​​from the acknowledgment messages, and output the ECN intensity and normalized RTT gradient of the current period. The scene feature extraction and recognition module is used to accumulate RTT measurement values ​​and ECN label information from multiple observation periods and extract observable feature vectors. The observable feature vectors are then input into the CART decision tree for scene discrimination, and the dominant scene category of the current network is output. The parameter scheduling and smoothing update module is used to select the corresponding target parameters from the preset parameter set according to the dominant scenario category, perform exponential weighted moving average smoothing on the target parameters, and output the currently effective energy leakage factor and rollback activation threshold. The congestion energy integration module is used to receive the ECN intensity and normalized RTT gradient output by the multidimensional signal observation module, as well as the energy leakage factor output by the parameter scheduling and smoothing update module. The energy leakage factor is used as the historical state retention weight to perform weighted fusion of the current period's ECN intensity and normalized RTT positive gradient, and output the congestion energy state. The dual-mode control module is used to receive the congestion energy state output by the congestion energy integration module and the back-off activation threshold output by the parameter scheduling and smooth update module, respectively map the congestion energy state to the smooth growth gain and the adaptive back-off modulation factor, and select the smooth growth gain or the adaptive back-off modulation factor to update the current congestion window according to whether there is an ECN flag in the current period, and output the updated congestion window. The closed-loop feedback module is used to apply the updated congestion window output by the dual-mode control module to the transmission of subsequent data packets and to trigger the scene feature extraction and recognition module to enter the next observation cycle.

Images