Adaptive low-latency live streaming method and system for sports

Abstract

The application discloses a self-adaptive low-delay sports scene live broadcast method and system, and particularly relates to the technical field of scene live broadcast. Through unified mapping and abnormality suppression of multi-source time sequence data, combined with a multi-scale sliding window predictor and a short-time autoregression and long-time trend perception algorithm, a more accurate bandwidth prediction result with interval confidence description is generated. Residual trajectory is used to extract bias accumulation, fluctuation intensity, error retention time and other features, and is synchronized with parameter switching log to analyze threshold crossing frequency, switching amplitude, direction alternation rate and critical zone retention time, forming critical oscillation characteristics, so that risk identification is more accurate. Through normalization and time sequence risk identifier calculation system, high-frequency oscillation risk score is calculated, and the risk evaluation result is mapped into an executable stabilization intervention strategy. The system state observation sequence after adjustment and execution is short-time evaluated and a feedback package is written back to form a closed loop, realizing online optimization of bandwidth predictor weight, decision threshold and cooling time.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of live streaming, more particularly, the present application relates to a self-adaptive low-latency live streaming method and system for sports scenes. BACKGROUND

[0002] In the self-adaptive low-latency live streaming for sports scenes, the system usually relies on bandwidth prediction results to dynamically adjust core parameters such as code rate, GOP length and redundancy rate, in order to strike a balance between picture quality and latency. However, in actual operation, the network bandwidth of sports scenes often presents high-frequency and unpredictable fluctuations. When the update frequency or accuracy of the bandwidth prediction model is insufficient, the predicted value will lag behind the real network state, forming bandwidth prediction error accumulation. This error, after entering the adaptive control link, will make the parameter adjustment logic frequently cross the set threshold, thus triggering multiple rounds of code rate, GOP and redundancy rate switching. Since these switching points are often concentrated in the threshold vicinity, the system is prone to fall into a "critical oscillation" state - that is, repeatedly performing the up and down operations in a short period of time. The bandwidth prediction error and the critical oscillation effect amplify each other, ultimately leading to the live streaming link constantly swinging at high frequency between picture quality and latency control, and failing to stably maintain in the expected low-latency and high-quality target interval, resulting in a significant decline in audience experience. SUMMARY

[0003] In order to overcome the above-mentioned defects of the prior art, the embodiments of the present application provide a self-adaptive low-latency live streaming method and system for sports scenes to solve the problems raised in the background art.

[0004] To achieve the above object, the present application provides the following technical scheme:

[0005] The self-adaptive low-latency live streaming method for sports scenes comprises the following steps:

[0006] Collecting multi-source operation data to generate a multi-source time series data stream, performing unified time base mapping and abnormality suppression processing on the multi-source time series data stream to generate a time series consistent dataset, applying a multi-scale sliding window predictor based on the time series consistent dataset, combining a short-term autoregression and a long-term trend perception algorithm to output a bandwidth prediction value and an interval confidence description, and forming a bandwidth prediction set;

[0007] Calculating a residual trajectory using the bandwidth prediction set and actual bandwidth samples, and extracting a bias accumulation, fluctuation intensity and error lag time feature in the residual trajectory to generate bandwidth prediction error accumulation information;

[0008] Time-synchronizing and aligning the bandwidth prediction error accumulation information and the adaptive parameter switching log, counting the threshold crossing frequency, switching amplitude, direction alternation rate and critical zone residence time, and generating adaptive parameter switching critical oscillation features;

[0009] The bandwidth prediction error accumulation information is normalized with the adaptive parameter switching critical oscillation feature, and is input into a time sequence risk identifier, to output a risk score and a risk level of the system in a high-frequency oscillation state of the code rate and delay strategy, and to generate a high-frequency oscillation risk assessment result;

[0010] The high-frequency oscillation risk assessment result is mapped into a stabilization intervention strategy token, the strategy token is converted into an executable parameter adjustment instruction, and the instruction is injected into a coding control link and a transmission scheduling link by using a delay-sensitive smoothing scheduling algorithm, to generate a system state observation sequence after adjustment execution;

[0011] The system state observation sequence after adjustment execution is short-time evaluated, an intervention effect feedback package is generated and written back to a risk control closed loop.

[0012] In a preferred embodiment, the multi-source time sequence data stream is subjected to unified time base mapping and abnormality suppression processing, to generate a time sequence consistent data set, specifically as follows:

[0013] The multi-source original time sequence data stream is subjected to field normalization and metadata injection, to generate a normalized time sequence data stream;

[0014] The normalized time sequence data stream is subjected to time synchronization correction and sampling reconstruction, to generate a synchronized time sequence data stream;

[0015] The synchronized time sequence data stream is subjected to abnormality detection and adaptive filtering, to form a time sequence consistent data set.

[0016] In a preferred embodiment, a multi-scale sliding window predictor is applied based on the time sequence consistent data set, combined with a short-time autoregressive and long-time trend perception algorithm, to output a bandwidth prediction value and an interval confidence description, to form a bandwidth prediction set, specifically as follows:

[0017] Multi-scale feature extraction and exogenous quantity fusion are performed on the time sequence consistent data set: instantaneous bandwidth gradient, instantaneous throughput volatility rate, and short-term packet loss rate are extracted in a short window; trend slope, periodic baseline, and seasonal decomposition are extracted in a medium / long window; and exogenous features are incorporated; to generate a multi-scale feature set;

[0018] The multi-scale feature set is input into a multi-scale sliding window predictor, a short-time autoregressive predictor, a fast online regressor, and a long-time trend perceiver are run in parallel according to a dynamic window strategy, and each window prediction result is adaptively weighted according to recent residual performance, to form a bandwidth prediction candidate set.

[0019] In a preferred embodiment, a residual trajectory is calculated using the bandwidth prediction set and an actual bandwidth sample, and bias accumulation, volatility intensity, and error retention time features are extracted in the residual trajectory, to generate bandwidth prediction error accumulation information, specifically as follows:

[0020] The time-aligned bandwidth prediction and the actual observed bandwidth sample calculate point-by-point residual, and generate an initial residual trajectory;

[0021] The initial residual trajectory is subjected to multi-scale denoising and detrending processing to generate a net residual trajectory;

[0022] The net residual trajectory is subjected to change point detection and segmented steady state division to generate a segmented residual sequence and segment boundary timestamp;

[0023] In the segmented residual sequence, the intra-segment residual statistics are calculated for each segment and accumulated by time to form a bias accumulation curve, and the segment-level metadata is recorded;

[0024] The bias accumulation curve is subjected to long-term offset identification and rate estimation to generate long-term offset identification and rate estimation;

[0025] The long-term offset identification and rate estimation are combined with the segmented residual sequence to calculate the fluctuation intensity time sequence by an adaptive sliding window, to generate the fluctuation intensity time sequence and label it with uncertainty;

[0026] The fluctuation intensity time sequence is subjected to peak / clustering detection and the amplitude, rising / falling slope and duration of each peak are counted to generate an error retention time feature list;

[0027] The error retention time feature list and the bias accumulation curve are subjected to time sequence coupling analysis to generate an error retention and bias accumulation coupling feature set, and it is packaged as a bandwidth prediction error accumulation information package, which includes: bias accumulation curve, fluctuation intensity sequence, error retention time feature list, error retention and bias accumulation coupling feature set.

[0028] In a preferred embodiment, the bandwidth prediction error accumulation information is time-synchronized with the adaptive parameter switching log, the threshold crossing frequency, switching amplitude, direction alternation rate and critical zone retention duration are counted to generate adaptive parameter switching critical oscillation features, as follows:

[0029] The bandwidth prediction error accumulation information and the adaptive parameter switching log are baseline-aligned according to their original timestamps to generate a time-aligned observation sequence;

[0030] The time-aligned observation sequence is subjected to resampling and time base unification, and the resampled segments are subjected to quality labeling and missing segment labeling to generate a synchronized, quality-labeled observation sequence;

[0031] The synchronized, quality-labeled observation sequence is subjected to robust anomaly suppression and pulse preservation, and confidence labels are added to the suppressed / preserved points to generate a cleaned, confidence information-labeled observation time sequence;

[0032] Converting the cleaned observation time series stream with confidence metadata into event stream: detecting and labeling each threshold crossing, switching start / end time and parameter values before and after switching, generating switching event list with before-after value labels;

[0033] Windowing and pairing the switching event with before-after value labels with the cleaned residual context, generating event-residual context tuples;

[0034] Calculating and appending switching amplitude indicators to the event-residual context tuples, generating event tuples with amplitude labels;

[0035] Direction labeling and direction sequence construction for the event tuples with amplitude labels: identifying the up / down alternation of consecutive event sequences, and calculating the raw count of direction alternation and the alternation rate weighted by amplitude on the sequence, generating the direction alternation rate measure;

[0036] Combining the direction alternation rate measure with the event tuples to determine the critical region membership: labeling the critical region entry / exit points according to the parameter threshold and pre-set hysteresis band, and merging consecutive critical region segments to calculate the critical region residence duration, generating the critical region event list;

[0037] Summarizing the critical oscillation feature set of adaptive parameter switching, including: threshold crossing frequency, normalized average switching amplitude, amplitude-weighted direction alternation rate, and average critical region residence duration; and attaching quality labels and time window identifiers to each item, forming a critical oscillation feature package that can be directly input to the downstream risk evaluator.

[0038] In a preferred embodiment, the bandwidth prediction error accumulation information and the critical oscillation features of adaptive parameter switching are normalized and input into a time series risk identifier, outputting the risk score and level of the system in the high-frequency oscillation state of the code rate and delay strategy, generating the high-frequency oscillation risk assessment result;

[0039] The risk score of the system in the high-frequency oscillation state of the code rate and delay strategy is obtained by weighted summation of the bandwidth prediction error accumulation information and the critical oscillation features of adaptive parameter switching.

[0040] In a preferred embodiment, the adaptive low-latency sports live streaming system includes a multi-source time series prediction processing module, a bandwidth prediction error analysis module, a critical oscillation analysis module of adaptive parameter switching, a high-frequency oscillation risk assessment module, a stabilization intervention strategy execution module, an intervention effect feedback and closed-loop update module;

[0041] A multi-source time series prediction processing module is configured to collect multi-source operation data, generate multi-source time series data streams, perform unified time base mapping and abnormality suppression processing on the multi-source time series data streams, generate time series consistent data sets, apply a multi-scale sliding window predictor based on the time series consistent data sets, combine a short-term autoregressive algorithm and a long-term trend perception algorithm, output bandwidth prediction values and interval confidence descriptions, and form a bandwidth prediction set;

[0042] A bandwidth prediction error analysis module is configured to calculate residual trajectories by using the bandwidth prediction set and actual bandwidth samples, extract bias accumulations, fluctuation intensities, and error residence time characteristics in the residual trajectories, and generate bandwidth prediction error accumulation information;

[0043] An adaptive parameter switching critical oscillation analysis module is configured to perform time synchronization alignment on the bandwidth prediction error accumulation information and adaptive parameter switching logs, count threshold crossing frequencies, switching amplitudes, direction alternation rates, and critical zone residence time lengths, and generate adaptive parameter switching critical oscillation characteristics;

[0044] A high-frequency oscillation risk assessment module is configured to perform normalization processing on the bandwidth prediction error accumulation information and adaptive parameter switching critical oscillation characteristics, input the time series risk identifier, and output risk scores and levels of the system being in a high-frequency oscillation state of the rate and delay strategy, and generate high-frequency oscillation risk assessment results;

[0045] A stabilization intervention strategy execution module is configured to map the high-frequency oscillation risk assessment results into stabilization intervention strategy tokens, convert the strategy tokens into executable parameter adjustment instructions, and inject the instructions into the encoding control link and the transmission scheduling link by using a delay-sensitive smoothing scheduling algorithm, and generate a system state observation sequence after adjustment execution;

[0046] An intervention effect feedback and closed-loop update module is configured to perform short-term evaluation on the system state observation sequence after adjustment execution, generate an intervention effect feedback package, and write back to the risk control closed loop.

[0047] The technical effects and advantages of the present application are as follows:

[0048] 1、The application realizes the dynamic regulation and control of key parameters such as code rate, GOP length and redundancy rate in the live link of the motion scene by fusing bandwidth prediction error accumulation analysis and adaptive parameter switching critical oscillation identification, can effectively inhibit the occurrence of high-frequency oscillation state, make the system still maintain the balance of picture quality and delay under the condition of network bandwidth fluctuation, through the unified mapping and abnormal suppression of multi-source time sequence data, combined with multi-scale sliding window predictor and short-term autoregressive and long-term trend perception algorithm, generate more accurate bandwidth prediction results with interval confidence description, thereby significantly reduce the prediction error accumulation, at the same time, the residual trajectory is used to extract bias accumulation, fluctuation intensity and error retention time, and the threshold crossing frequency, switching amplitude, direction alternation rate and critical zone retention time are analyzed synchronously with the parameter switching log, the critical oscillation characteristics are formed, the risk identification is more accurate, the high-frequency oscillation risk score of the system is calculated through normalization and time sequence risk identifier, and the risk evaluation result is mapped into executable stabilization intervention strategy, then the delay sensitive smoothing scheduling injection coding and transmission control link can smoothly transition parameter changes in the adjustment execution process, avoid picture jitter or delay surge caused by mutation, finally, the system state observation sequence after adjustment execution is evaluated for a short time and feedback package is written back to form a closed loop, online optimization of bandwidth predictor weight, decision threshold and cooling time is realized.

[0049] 2、The application can significantly improve the stability and reliability of the live system in the motion scene, reduce the high-frequency fluctuation risk of picture quality and delay, enable the audience to obtain continuous high-quality and low-delay viewing experience, and improve the adaptability and self-regulation efficiency of the system to complex network environment. BRIEF DESCRIPTION OF DRAWINGS

[0050] In order to facilitate those skilled in the art to understand, the application will be further described below with reference to the drawings;

[0051] Figure 1 The flowchart of the method of the embodiment 1 of the application is shown in the figure;

[0052] Figure 2 The flowchart of the system of the embodiment 2 of the application is shown in the figure. DETAILED DESCRIPTION

[0053] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor belong to the protection scope of the application.

[0054] Embodiment 1: Figure 1The adaptive low-delay live broadcast method for a moving scene according to the present application comprises the following steps:

[0055] Collecting multi-source operation data to generate a multi-source time sequence data stream, performing unified time base mapping and abnormality suppression processing on the multi-source time sequence data stream to generate a time sequence consistent data set, applying a multi-scale sliding window predictor based on the time sequence consistent data set, combining a short-term autoregressive algorithm and a long-term trend perception algorithm to output a bandwidth prediction value and an interval confidence description, and forming a bandwidth prediction set;

[0056] Calculating a residual trajectory using the bandwidth prediction set and actual bandwidth samples, and extracting a bias accumulation, fluctuation intensity and error residence time feature in the residual trajectory to generate bandwidth prediction error accumulation information;

[0057] Synchronizing and aligning the bandwidth prediction error accumulation information and an adaptive parameter switching log in time, and statistically calculating a threshold crossing frequency, a switching amplitude, a direction alternation rate and a critical zone residence time to generate adaptive parameter switching critical oscillation features;

[0058] Performing normalization processing on the bandwidth prediction error accumulation information and the adaptive parameter switching critical oscillation features, and inputting them into a time sequence risk identifier to output a risk score and a level of a system in a high-frequency oscillation state of a code rate and a delay strategy, and generating a high-frequency oscillation risk assessment result;

[0059] Mapping the high-frequency oscillation risk assessment result into a stabilization intervention strategy token, converting the strategy token into an executable parameter adjustment instruction, and using a delay-sensitive smoothing scheduling algorithm to inject the instruction into a coding control link and a transmission scheduling link to generate a system state observation sequence after adjustment execution;

[0060] Performing short-term evaluation on the system state observation sequence after adjustment execution to generate an intervention effect feedback package and write it back to a risk control closed loop;

[0061] Collecting multi-source operation data (including a coding end output code rate and a frame motion vector distribution, a transmission link instantaneous bandwidth sample, a round-trip delay / one-way delay, a packet loss / retransmission count, a receiving end buffer occupancy and a decoding delay, etc.) to generate a multi-source time sequence data stream.

[0062] Performing unified time base mapping and abnormality suppression processing on the multi-source time sequence data stream to generate a time sequence consistent data set, specifically as follows:

[0063] Performing field standardization and metadata injection (adding sampling accuracy, timestamp source identification, path label, device identification and sample quality mark) on the multi-source original time sequence data stream to generate a standardized time sequence data stream;

[0064] Time synchronization correction and sampling reconstruction are performed on the normalized time series data stream (clock drift calibration, time base alignment, interpolation and downsampling if necessary, and missing segment tagging are performed, and the interpolation confidence is recorded), to generate a synchronized time series data stream;

[0065] Anomaly detection and adaptive filtering are performed on the synchronized time series data stream (isolated point identification, short-time jump suppression, burst reservation strategy and local smoothing are performed, the sampling rate is increased to reserve key burst signals according to event triggers, and quality confidence labels are attached to each segment of data), to form a time-consistent data set;

[0066] A multi-scale sliding window predictor is applied based on the time-consistent data set, combining short-term autoregression and long-term trend perception algorithm, to output bandwidth prediction values and interval confidence description, forming a bandwidth prediction set, as follows:

[0067] Multi-scale feature extraction and exogenous quantity fusion are performed on the time-consistent data set: instantaneous bandwidth gradient, instantaneous throughput volatility rate, and short-term packet loss rate are extracted in a short window; trend slope, periodic baseline, and seasonal decomposition are extracted in a medium / long window; and exogenous features such as motion complexity and view switching events are integrated; a multi-scale feature set is generated;

[0068] The multi-scale feature set is input into a multi-scale sliding window predictor, which runs a short-term autoregressive predictor, a fast online regressor, and a long-term trend perceiver in parallel according to a dynamic window strategy (mutation / gear shifting detection is performed in parallel to identify state mutations), and adaptively weights the prediction results of each window according to the recent residual performance, to form a bandwidth prediction candidate set;

[0069] Uncertainty estimation and interval confidence description are performed on the bandwidth prediction candidate set (quantile intervals are generated using model set distribution, bootstrap sampling, or approximate Bayesian posterior estimation, and online calibration is performed using historical coverage rate), to generate a bandwidth prediction set and corresponding interval confidence description and confidence label;

[0070] Confidence applicability verification and meta-information annotation are performed on the bandwidth prediction set (confidence interval coverage rate is verified based on recent residual statistics, conservative predictor is triggered in high uncertainty window, and predictor priority and applicable time window are recorded), to output the final bandwidth prediction set and confidence meta-information available for downstream use, and to record the prediction and actual observation in parallel for subsequent online reweighting and offline retraining.

[0071] Residual trajectory is calculated using the bandwidth prediction set and actual bandwidth samples, and bias accumulation, volatility intensity, and error residence time features are extracted from the residual trajectory to generate bandwidth prediction error accumulation information, as follows:

[0072] calculating point-wise residuals (at time point t, the point-wise residual r(t) is equal to the observed value minus the predicted value) from the time-aligned bandwidth prediction and the actual observed bandwidth samples;

[0073] performing multi-scale denoising and detrending on the initial residual trajectory (short-window noise suppression, medium-long term baseline separation, preserving abrupt pulses), generating a net residual trajectory;

[0074] performing change point detection and piecewise stationary segmentation on the net residual trajectory (using online change point detection and minimum description length criteria), generating a segmented residual sequence and segment boundary timestamps;

[0075] calculating intra-segment residual statistics and accumulating into a bias accumulation curve, and recording segment-level metadata (including intra-segment mean accumulation, inter-segment offset direction and accumulation duration) for each segment in the segmented residual sequence;

[0076] performing long-term offset identification and rate estimation on the bias accumulation curve (identifying long-term drift based on inter-segment trend consistency and accumulation slope), generating long-term offset identification and rate estimation;

[0077] the performing long-term offset identification and rate estimation on the bias accumulation curve comprises:

[0078] preliminarily smoothing the bias accumulation curve by a pre-set minimum segment length to suppress high-frequency noise, generating a smoothed bias curve;

[0079] performing seasonal and periodic decomposition on the smoothed bias curve (extracting baseline trend, periodic component and residual component), generating a baseline trend sequence and periodic component annotation;

[0080] re-segmenting the baseline trend sequence by previous segment boundaries, generating a segmented baseline sequence and start-end timestamps for each segment;

[0081] calculating robust slope estimation (using anomaly-resistant regression methods such as Theil-Sen or robust least squares) for each segment of the segmented baseline sequence, generating a segment-level slope set and segment-level slope confidence interval;

[0082] performing direction consistency detection on the segment-level slope set (judging whether the majority of segment slopes are consistent in sign), and calculating slope sign proportion and adjacent segment slope sign transition rate, generating an inter-segment trend consistency measure (the trend consistency measure is obtained by weighted summation of the slope sign proportion and the adjacent segment slope sign transition rate);

[0083] accumulating the segment-level slope set into a cumulative slope curve by time weighting (giving higher weight to recent segments or weighting according to segment length / energy);

[0084] Perform significance test on the cumulative slope curve (using non-parametric trend test or bootstrap method to assess the significance of cumulative slope), generate long-term bias significance indicator and confidence interval;

[0085] Determine long-term bias flag by combining long-term bias significance indicator and inter-segment trend consistency measure: set long-term bias flag as "present" when significance passes and trend consistency exceeds threshold, otherwise set as "not detected"; generate long-term bias flag;

[0086] Perform rate estimation on the cumulative slope curve in the case of long-term bias confirmed as "present" (the rate estimation is obtained by weighted average of segment slopes);

[0087] Calculate volatility intensity time series by adaptive sliding window on long-term bias flag, rate estimation and segmented residual series (window length is adaptively adjusted according to historical volatility, and the indicator is estimated by robust variance / quartile range or energy spectrum), generate volatility intensity time series and label uncertainty;

[0088] Perform peak / clustering detection on volatility intensity time series and count the amplitude, rising / falling slope and duration of each peak, generate error retention time feature list (record peak start and end, peak energy and duration window);

[0089] Perform time series coupling analysis on error retention time feature list and bias accumulation curve (quantified by Pearson correlation coefficient), generate error retention and bias accumulation coupling feature set (including Pearson correlation coefficient, relative time sequence), and package it as bandwidth prediction error accumulation information package, which includes: bias accumulation curve, volatility intensity sequence, error retention time feature list, error retention and bias accumulation coupling feature set.

[0090] Synchronize bandwidth prediction error accumulation information and adaptive parameter switching log by time, count threshold crossing frequency, switching amplitude, direction alternation rate and critical zone retention duration, generate adaptive parameter switching critical oscillation features, as follows:

[0091] Baseline align bandwidth prediction error accumulation information and adaptive parameter switching log according to their original time stamps, generate time-aligned observation sequence;

[0092] Resample and time-base unify the time-aligned observation sequence (using the required minimum resolution or dynamic time warping calibration), and perform quality labeling and missing segment labeling on the resampled segments, generate synchronized, quality-labeled observation sequence;

[0093] Robust outlier suppression and pulse retention (using median filter combined with event retention strategy) on synchronized, quality-tagged observation sequence, and appending confidence labels at suppression / retention points, generating cleaned, confidence-tagged observation time series stream;

[0094] Converting cleaned, confidence-tagged observation time series stream to event stream: detecting and labeling each threshold crossing (including up and down crossing directions), switching start / end time points, and pre / post switching parameter values, generating a list of switching events with pre / post value labels;

[0095] Windowing pairing switching events with pre / post value labels and cleaned residual context (adaptive window length adjusted according to local fluctuation intensity, window containing short-term baseline before switching and short-term response after switching), generating event-residual context tuples;

[0096] It should be noted that the cleaned residual context refers to the local time period data retained after outlier suppression, noise removal and trend separation from the bandwidth prediction residual trajectory. This time period is centered on a specific event (such as adaptive parameter switching), and contains residual values, trend direction, fluctuation intensity, etc. before and after the event. It not only contains the numerical sequence of residual, but also retains the quality labels (such as "trusted after suppression" / "partially missing") and processing meta-information (denoising method, filter window length, etc.) related to this time period. The purpose is to provide a clean and contextually meaningful reference window when analyzing the relationship between switching events and residual changes, preventing original noise from interfering with judgment.

[0097] Calculating and appending switching amplitude indicators (absolute amplitude and normalized amplitude relative to baseline) to event-residual context tuples, generating event tuples with amplitude labels;

[0098] Directional labeling and direction sequence construction on event tuples with amplitude labels: identifying rising / falling alternation of continuous event sequence, and calculating original count of direction alternation and alternation rate weighted by amplitude on the sequence, generating direction alternation rate measure;

[0099] Combining direction alternation rate measure with event tuples to determine critical zone membership: labeling critical zone entry / exit points according to parameter threshold and pre-set hysteresis band, and merging consecutive critical zone segments to calculate critical zone residence time, generating critical zone event list;

[0100] It should be noted that the hysteresis band refers to the double threshold interval or buffer interval in adaptive switching to prevent parameters from repeatedly switching back and forth due to slight fluctuations; including the threshold value for entering the critical region (triggering switching) and the threshold value for exiting the critical region (deactivating switching), there is a certain interval (bandwidth or delay value difference) between the two threshold values, forming a buffer area in the "bandwidth-delay space", within this interval, even if the monitoring value fluctuates back and forth, switching will not be triggered immediately, thereby reducing the probability of "jitter" or "critical oscillation". The identification of the hysteresis band is to determine the residence time of the system in the critical region, so as to evaluate whether it is in a high-risk oscillation state.

[0101] The direction alternation rate metric is combined with the event tuple to determine critical region membership, and a critical region event list is generated, as follows:

[0102] The direction alternation rate metric is combined with the event tuple to determine critical region membership, and a critical region event list is generated, as follows:

[0103] The direction alternation rate metric is combined with the event tuple to determine critical region membership, and a critical region event list is generated, as follows:

[0104] It should be noted that the event parameter value refers to the specific value of the core operating parameter adjusted in the system before and after a certain adaptive parameter switching event occurs.

[0105] For example:

[0106] Encoding related parameters: video encoding rate (e.g. 3.2 Mbps changes to 2.5 Mbps); GOP length (e.g. 60 frames changes to 30 frames); encoding quantization parameter QP value (e.g. 28 changes to 32); encoding resolution (e.g. 1920x1080 changes to 1280x720);

[0107] Transmission related parameters: sending end target throughput (e.g. 5 Mbps changes to 4 Mbps); FEC redundancy ratio (e.g. 10% changes to 15%); adaptive buffer target length (e.g. 300 ms changes to 250 ms);

[0108] Play related parameters (if bidirectional adaptive): play buffer target delay (e.g. 200 ms changes to 180 ms); play end decoding frame rate limit (e.g. 60 fps changes to 48 fps).

[0109] The event-by-event critical state flag is tracked by state transition (entry change to maintenance change to exit), and the upper and lower boundaries of the hysteresis band are used to suppress frequent state jitter caused by measurement noise, to generate a noise-suppressed critical state time series;

[0110] The noise-suppressed critical state time series is merged into continuous segments (merge rule: adjacent state segments with a gap less than the minimum gap threshold are merged into the same segment), to generate a list of critical region candidate segments;

[0111] The critical region candidate segment list is cross-screened with the direction alternation pattern of the event sequence (keeping the segments with a direction alternation rate higher than the set percentage threshold), to generate a filtered critical region effective segment set;

[0112] The critical region effective segment set is calculated for each segment's residence time (segment end time minus segment start time), and the statistical quantities: maximum residence time, average residence time, residence time variance are attached, to generate a critical region segment residence feature set;

[0113] The critical region segment residence feature set is jointly labeled with the alternation rate and switching amplitude distribution within the segment (generate a multi-dimensional segment image, including residence intensity, direction alternation activity, and amplitude fluctuation intensity), to form a structured critical region event list as a direct input for downstream risk identification and strategy generation;

[0114] The adaptive parameter switching critical oscillation feature set is summarized, including: threshold crossing frequency, normalized average switching amplitude, amplitude-weighted direction alternation rate, and average critical region residence time; and for each item, a quality label and a time window identifier are attached, to form a critical oscillation feature package that can be directly input into the downstream risk evaluator.

[0115] The bandwidth prediction error accumulation information and the adaptive parameter switching critical oscillation features are normalized and input into the time series risk identifier, to output the risk score and level of the system in the high-frequency oscillation state of the code rate and delay strategy, to generate a high-frequency oscillation risk assessment result;

[0116] The risk score of the system in the high-frequency oscillation state of the code rate and delay strategy is obtained by weighted summation of the bandwidth prediction error accumulation information and the adaptive parameter switching critical oscillation features.

[0117] The risk score of the system in the high-frequency oscillation state of the code rate and delay strategy is compared with the preset score stage threshold, to classify the high-frequency oscillation risk;

[0118] The classification result and the risk score are combined to form a high-frequency oscillation risk assessment result.

[0119] The high-frequency oscillation risk assessment result is mapped to a stabilization intervention strategy token, the strategy token is converted into an executable parameter adjustment instruction, and the instruction is injected into the encoding control link and the transmission scheduling link using a delay-sensitive smoothing scheduling algorithm to generate a system state observation sequence after adjustment execution, as follows:

[0120] The high-frequency oscillation risk assessment result is mapped to a set of strategy tokens, each token containing (risk level, recommended measure list, impact parameter set, action duration upper limit and safety boundary), and a strategy token set is generated;

[0121] The strategy token set is prioritized according to risk level, current system SLA and business priority, and each token is supplemented with parameter constraints (maximum allowed amplitude, minimum residence time, mutual exclusion rules) based on the safety boundary library and the dependency graph, generating a constrained strategy token queue;

[0122] The constrained strategy token queue is matched and mapped with the pre-defined parameter template library to generate a parameter adjustment draft set, which includes recommended values and execution windows for the encoding side (target bit rate upper limit, QP change, GOP / key frame interval, key frame priority, redundancy ratio) and the transmission side (packet scheduling priority, fragmentation / retransmission strategy, FEC ratio, path selection recommendation, distribution rate upper limit);

[0123] The parameter adjustment draft set is subjected to constraint checking (boundary overrun detection, mutual exclusion conflict resolution, rollback path confirmation), and when conflicts occur, the conflicts are resolved according to priority and dependency (such as replacing delay-sensitive measures with degraded image measures), generating an operation instruction set with constraints and rollback strategies;

[0124] The operation instruction set is divided into time segments according to instruction priority, impact range and minimum interruption principle, and each instruction is assigned an execution start and end window, a maximum execution slope (maximum parameter change rate) and a command frequency limit (token bucket style), forming a scheduling time table to be injected;

[0125] The delay-sensitive smoothing scheduling algorithm is applied to the scheduling time table, which uses (priority, deadline, system delay budget, maximum impact budget, minimum residence time) as scheduling constraints, generates a gradual parameter curve using piecewise linear / smooth S (sigmoid) slope or exponential decay function, and reserves a short preemption window in front of high-confidence / high-priority instructions, outputs a gradual execution plan and generates a command package with version number;

[0126] Convert the progressive execution plan into a series of delivery commands recognizable by the encoding control (such as setting the instantaneous code rate upper limit, submitting QP adjustment, key frame insertion order, adjusting the FEC ratio, setting the transmission queue priority), and use atomic or phased commit strategies (including pre-commit-confirmation-implementation three-stage or two-phase commit) to deliver the command package, generate execution confirmation and initial response indication;

[0127] Upon execution confirmation and initial response indication triggering, high-frequency sampling of the system is performed according to the sampler, key observation quantities (instantaneous code rate output, actual GOP / key frame time, end-to-end delay distribution, packet loss / retransmission statistics, FEC effectiveness, buffer occupancy, switching event count, etc.) are collected and recorded, and the system state observation sequence after adjustment execution is formed in time sequence.

[0128] The system state observation sequence after adjustment execution is evaluated for a short time, and the intervention effect feedback package is generated and written back to the risk control closed loop, as follows:

[0129] The system state observation sequence after adjustment execution is evaluated for a short time (compare expected and actual delay / picture quality / stability indicators), if the observation sequence triggers rollback conditions (such as delay exceeding budget, jitter amplification or violation of key service SLA), then roll back to the previous confirmable state according to the rollback strategy in the instruction package and generate rollback record; otherwise, confirm the implementation and generate the final execution state;

[0130] The final execution state and observation sequence generate intervention effect feedback package and write back to the risk control closed loop: the final execution state, observation sequence summary, instruction execution log and rollback record are packaged into intervention effect feedback package, and are written back to the prediction / risk assessment and strategy priority maintenance process according to time window and quality label,

[0131] Versioned archive and security audit labeling of intervention effect feedback package and corresponding strategy token, operation instruction set and execution log are performed to form a traceable strategy execution history record.

[0132] The application realizes dynamic regulation and control of key parameters such as code rate, GOP length and redundancy rate in the live scene link by fusing bandwidth prediction error accumulation analysis and adaptive parameter switching critical oscillation identification, can effectively inhibit the occurrence of high-frequency oscillation state, make the system still maintain the balance of picture quality and delay under the condition of network bandwidth fluctuation, through unified mapping and abnormal suppression of multi-source time sequence data, combined with multi-scale sliding window predictor and short-term autoregressive and long-term trend perception algorithm, generate more accurate bandwidth prediction results with interval confidence description, thereby significantly reduce the prediction error accumulation, at the same time, the residual trajectory is used to extract the bias accumulation, fluctuation intensity and error retention time, and the threshold crossing frequency, switching amplitude, direction alternation rate and critical zone retention time are analyzed synchronously with the parameter switching log, the critical oscillation characteristics are formed, the risk identification is more accurate, the high-frequency oscillation risk score is calculated by normalization and time sequence risk identifier, and the risk evaluation result is mapped into an executable stabilization intervention strategy, and then the delay sensitive smoothing scheduling injection coding and transmission control link can smoothly transition the parameter change in the adjustment execution process, avoid the picture jitter or delay surge caused by mutation, finally, the system state observation sequence after adjustment execution is evaluated for a short time and a feedback package is written back to form a closed loop, realizing online optimization of bandwidth predictor weight, decision threshold and cooling time.

[0133] The application can significantly improve the stability and reliability of the live system in the motion scene, reduce the high-frequency fluctuation risk of picture quality and delay, and enable the audience to obtain a continuous high-quality and low-delay viewing experience, while improving the adaptability and self-regulation efficiency of the system to complex network environment.

[0134] Embodiment 2: This embodiment is an introduction to the adaptive low-delay live system in the motion scene, as shown in Figure 2 The adaptive low-delay live system in the motion scene includes a multi-source time sequence prediction processing module, a bandwidth prediction error analysis module, an adaptive parameter switching critical oscillation analysis module, a high-frequency oscillation risk evaluation module, a stabilization intervention strategy execution module, and an intervention effect feedback and closed loop updating module.

[0135] The multi-source time sequence prediction processing module is used for collecting multi-source running data, generating multi-source time sequence data stream, performing unified time base mapping and abnormal suppression processing on the multi-source time sequence data stream, generating time sequence consistent data set, applying multi-scale sliding window predictor based on the time sequence consistent data set, combining short-term autoregressive and long-term trend perception algorithm, outputting bandwidth prediction value and interval confidence description, and forming bandwidth prediction set.

[0136] The bandwidth prediction error analysis module is used for calculating residual trajectory by using the bandwidth prediction set and actual bandwidth sample, and extracting bias accumulation, fluctuation intensity and error retention time features in the residual trajectory, and generating bandwidth prediction error accumulation information.

[0137] The adaptive parameter switching critical oscillation analysis module is configured to time-synchronize and align the bandwidth prediction error accumulation information and the adaptive parameter switching log, count the threshold crossing frequency, the switching amplitude, the direction alternation rate and the critical zone residence time, and generate adaptive parameter switching critical oscillation characteristics;

[0138] The high-frequency oscillation risk assessment module is configured to normalize the bandwidth prediction error accumulation information and the adaptive parameter switching critical oscillation characteristics, input the time sequence risk identifier, output the risk score and level of the system in the high-frequency oscillation state of the code rate and the delay strategy, and generate a high-frequency oscillation risk assessment result.

[0139] The stabilization intervention strategy execution module is configured to map the high-frequency oscillation risk assessment result to a stabilization intervention strategy token, convert the strategy token into an executable parameter adjustment instruction, and inject the instruction into the encoding control link and the transmission scheduling link by using a delay-sensitive smoothing scheduling algorithm, and generate a system state observation sequence after adjustment execution.

[0140] The intervention effect feedback and closed-loop update module is configured to perform short-time evaluation on the system state observation sequence after adjustment execution, generate an intervention effect feedback package and write it back to the risk control closed loop.

[0141] The above formulas are dimensionless numerical calculations, the formulas are obtained by software simulation of a large amount of data to obtain a formula of the latest real situation, and the preset parameters in the formula are set by a person skilled in the art according to the actual situation.

[0142] The above-described embodiments can be implemented in part or in whole through software, hardware, firmware or any combination thereof. When implemented in software, the above-described embodiments can be implemented in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When loaded and executed by a computer, the computer instructions or computer programs cause the computer to perform the processes or functions described in the embodiments of the present application. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable apparatus. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another computer-readable storage medium, for example, the computer instructions can be transmitted from one website site, computer, server or data center to another website site, computer, server or data center through a wired or wireless (such as infrared, wireless, microwave, etc.) manner. The computer-readable storage medium can be any available medium accessible by a computer or a data storage device such as a server, data center, etc. containing a set of one or more available media. The available media can be a magnetic medium (such as a floppy disk, a hard disk, a magnetic tape), an optical medium (such as a DVD), or a semiconductor medium. The semiconductor medium can be a solid state disk.

[0143] It should be understood that, in various embodiments of the present application, the size of the sequence number of each process described above does not mean the order of execution, and the execution order of each process should be determined according to its function and inherent logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0144] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the system and method described above can refer to the corresponding process in the foregoing method embodiments, which will not be described here.

[0145] In several embodiments provided in the present application, it should be understood that the disclosed system and method can be implemented in other ways.

[0146] The above is merely specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any skilled person in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. An adaptive low-latency motion scene live streaming method, characterized by: Includes the following steps: Collect multi-source operational data to generate multi-source time-series data streams. Perform unified time base mapping and anomaly suppression processing on the multi-source time-series data streams to generate a time-series consistent dataset. Apply a multi-scale sliding window predictor based on the time-series consistent dataset, and combine short-term autoregressive and long-term trend-aware algorithms to output bandwidth prediction values ​​and interval confidence descriptions to form a bandwidth prediction set. The short-term autoregressive and long-term trend-aware algorithm includes extracting instantaneous bandwidth gradient, instantaneous throughput volatility, and short-term packet loss rate within a short window. Extract trend slope, cycle baseline, and seasonality decomposition within a medium / long window; The residual trajectory is calculated using the bandwidth prediction set and the actual bandwidth sample, and the characteristics of cumulative bias, fluctuation intensity and error residence time are extracted from the residual trajectory to generate bandwidth prediction error accumulation information. The extraction logic of the offset cumulative amount is as follows: calculate the point-by-point residual between the time-aligned bandwidth prediction and the actual observed bandwidth sample, and generate the initial residual trajectory; Multi-scale denoising and detrending processing is performed on the initial residual trajectory to generate the net residual trajectory; Perform change point detection and segmented steady-state division on the net residual trajectory to generate segmented residual sequences and segment boundary timestamps; In the segmented residual sequence, the residual statistics within each segment are calculated and accumulated over time to form a bias cumulative curve; The bandwidth prediction error accumulation information is synchronized with the adaptive parameter switching log in time, and the threshold crossing frequency, switching amplitude, direction alternation rate and critical zone dwell time are statistically analyzed to generate adaptive parameter switching critical oscillation characteristics. The adaptive parameters are the event parameter values ​​of the core operating parameters that are adjusted in the system before and after the adaptive parameter switching event, including encoding parameters, transmission parameters, and playback parameters. The threshold crossing frequency is the frequency of events detected and labeled for each threshold crossing; The switching amplitude is the absolute amplitude and the normalized amplitude relative to the baseline; The direction alternation rate is used to identify the rising / falling alternation of a continuous event sequence, and to calculate the original count of direction alternation and the magnitude-weighted alternation rate on the sequence to generate a direction alternation rate metric. The bandwidth prediction error accumulation information and the critical oscillation characteristics of adaptive parameter switching are normalized and input into the time-series risk identifier. The output is the risk score and level of the system in the high-frequency oscillation state of the bit rate and delay strategy, and the high-frequency oscillation risk assessment result is generated. The high-frequency oscillation risk assessment results are mapped to stabilization intervention strategy tokens, the strategy tokens are converted into executable parameter adjustment instructions, and the instructions are injected into the coded control link and the transmission scheduling link using a delay-sensitive smooth scheduling algorithm to generate a system state observation sequence after the adjustment is executed. A short-term evaluation of the system status observation sequence after the adjustment is implemented is performed, an intervention effect feedback package is generated and written back to the risk control closed loop.

2. The adaptive low-latency motion scene live streaming method according to claim 1, characterized in that: A unified time base mapping and anomaly suppression process is performed on multi-source time-series data streams to generate a time-consistent dataset, as detailed below: Field normalization and metadata injection are performed on multi-source raw time-series data streams to generate normalized time-series data streams; Perform time synchronization correction and sampling reconstruction on the normalized time-series data stream to generate a synchronized time-series data stream; Anomaly detection and adaptive filtering are performed on the synchronized time-series data stream to form a time-consistent dataset.

3. The adaptive low-latency motion scene live streaming method according to claim 2, characterized in that: Based on the aforementioned time-consistent dataset, a multi-scale sliding window predictor is applied, combined with short-term autoregressive and long-term trend-aware algorithms, to output bandwidth prediction values ​​and interval confidence descriptions, forming a bandwidth prediction set, as detailed below: Multi-scale feature extraction and exogenous feature fusion are performed on the time-series consistent dataset: instantaneous bandwidth gradient, instantaneous throughput volatility, and short-term packet loss rate are extracted in the short window; trend slope, periodic baseline, and seasonal decomposition are extracted in the medium / long window; and exogenous features are incorporated. Generate multi-scale feature sets; The multi-scale feature set is input into the multi-scale sliding window predictor. The short-term autoregressive predictor, the fast online regressor, and the long-term trend perceptron are run in parallel according to the dynamic window strategy. The prediction results of each window are adaptively weighted according to the recent residual performance to form a bandwidth prediction candidate set.

4. The adaptive low-latency motion scene live streaming method according to claim 3, characterized in that: The residual trajectory is calculated using the bandwidth prediction set and the actual bandwidth sample. The cumulative bias, fluctuation intensity, and error residence time features are then extracted from the residual trajectory to generate bandwidth prediction error accumulation information, as detailed below: The pointwise residuals of the time-aligned bandwidth predictions and the actual observed bandwidth samples are calculated, and the initial residual trajectory is generated. Multi-scale denoising and detrending processing is performed on the initial residual trajectory to generate the net residual trajectory; Perform change point detection and segmented steady-state division on the net residual trajectory to generate segmented residual sequences and segment boundary timestamps; In the segmented residual sequence, the residual statistics within each segment are calculated and accumulated over time to form a bias cumulative curve, and segment-level metadata is recorded; Long-term offset identification and rate estimation are performed on the offset cumulative curve to generate long-term offset identifiers and rate estimates. The fluctuation intensity time series is calculated by combining the long-term migration identifier and rate estimation with the segmented residual sequence using an adaptive sliding window, generating the fluctuation intensity time series and labeling it with uncertainty. Peak / cluster detection is performed on the time series of fluctuation intensity, and the amplitude, rise / fall slope and duration of each peak are counted to generate a list of error residence time characteristics. The error residence time feature list and the bias accumulation curve are subjected to time-series coupling analysis to generate the error residence and bias accumulation coupling feature set, which is then packaged into a bandwidth prediction error accumulation information package. This information package includes: bias accumulation curve, fluctuation intensity sequence, error residence time feature list, and error residence and bias accumulation coupling feature set.

5. The adaptive low-latency motion scene live streaming method according to claim 4, characterized in that: The bandwidth prediction error accumulation information is synchronized with the adaptive parameter switching log in time, and the threshold crossing frequency, switching amplitude, direction alternation rate, and critical zone dwell time are statistically analyzed to generate adaptive parameter switching critical oscillation characteristics, as follows: The bandwidth prediction error accumulation information and the adaptive parameter switching log are baseline aligned according to their original timestamps to generate a time-aligned observation sequence. Resampling and time base unification are performed on the time-aligned observation sequence, and quality and missing segment markings are applied to the resampled segments to generate synchronized, quality-marked observation sequences. Robust anomaly suppression and impulse retention are performed on synchronized, quality-labeled observation sequences, and confidence labels are added at the suppression / retention points to generate a clean observation time series stream with confidence element information; Convert the cleaned observation time series with confidence element information into an event stream: detect and label each threshold crossing, the start / end time of the switch, and the parameter values ​​before and after the switch, and generate a list of switch events with labels of the before and after values; The switching events with before and after value labels are windowed and paired with the cleaned residual context to generate event-residual context tuples; The event-residual context tuple is calculated and a switching magnitude indicator is attached to generate an event tuple with magnitude annotation; The event tuples with amplitude labels are directionally labeled and a directional sequence is constructed: the rising / falling alternation of continuous event sequences is identified, and the original count of directional alternation and the alternation rate weighted by amplitude are calculated on the sequence to generate a directional alternation rate metric. The critical zone membership is determined by combining the direction alternation rate metric with event tuples: the entry / exit points of the critical zone are marked according to the parameter threshold and the preset hysteresis band, and continuous critical zone segments are merged to calculate the critical zone residence time and generate a critical zone event list. The adaptive parameter switching critical oscillation feature set is summarized, including: threshold crossing frequency, normalized average switching amplitude, amplitude-weighted direction alternation rate, and average critical zone dwell time. Each item is accompanied by a quality label and a time window identifier, forming a critical oscillation feature package that can be directly input into downstream risk assessors.

6. The adaptive low-latency motion scene live streaming method according to claim 5, characterized in that: The bandwidth prediction error accumulation information and the adaptive parameter switching critical oscillation characteristics are normalized and input into the timing risk identifier. The risk score and level of the system in the high-frequency oscillation state of the bit rate and delay strategy are output to generate a high-frequency oscillation risk assessment result, including: the risk score of the system in the high-frequency oscillation state of the bit rate and delay strategy is obtained by weighted summation of the bandwidth prediction error accumulation information and the adaptive parameter switching critical oscillation characteristics.

7. An adaptive low-latency motion scene live streaming system, used to implement the adaptive low-latency motion scene live streaming method according to any one of claims 1-6, characterized in that: It includes a multi-source time series prediction processing module, a bandwidth prediction error analysis module, an adaptive parameter switching critical oscillation analysis module, a high-frequency oscillation risk assessment module, a stabilization intervention strategy execution module, and an intervention effect feedback and closed-loop update module; The multi-source time series prediction processing module is used to collect multi-source running data, generate multi-source time series data streams, perform unified time base mapping and anomaly suppression processing on the multi-source time series data streams, generate a time series consistent dataset, apply a multi-scale sliding window predictor based on the time series consistent dataset, and combine short-time autoregressive and long-time trend perception algorithms to output bandwidth prediction values ​​and interval confidence descriptions to form a bandwidth prediction set. The bandwidth prediction error analysis module is used to calculate the residual trajectory using the bandwidth prediction set and the actual bandwidth sample, and extract the cumulative bias, fluctuation intensity and error residence time characteristics from the residual trajectory to generate bandwidth prediction error accumulation information. The adaptive parameter switching critical oscillation analysis module is used to synchronize and align bandwidth prediction error accumulation information with adaptive parameter switching logs in time, and to statistically analyze threshold crossing frequency, switching amplitude, direction alternation rate and critical zone dwell time to generate adaptive parameter switching critical oscillation characteristics. The high-frequency oscillation risk assessment module is used to normalize the bandwidth prediction error accumulation information and the critical oscillation characteristics of adaptive parameter switching, and input them into the timing risk identifier. It outputs the risk score and level of the system in the high-frequency oscillation state of the bit rate and delay strategy, and generates the high-frequency oscillation risk assessment result. The stabilization intervention strategy execution module is used to map the high-frequency oscillation risk assessment results into stabilization intervention strategy tokens, convert the strategy tokens into executable parameter adjustment instructions, and use a delay-sensitive smooth scheduling algorithm to inject the instructions into the coded control link and the transmission scheduling link to generate a system state observation sequence after the adjustment is executed. The intervention effect feedback and closed-loop update module is used to perform short-term evaluation of the system state observation sequence after the adjustment is implemented, generate an intervention effect feedback package, and write it back to the risk control closed loop.

Images