Bandwidth control and visitor traffic routing optimization method and system for multi-scenario live broadcast

Abstract

The present invention provides a bandwidth control and visitor traffic routing optimization method and system for multi-scenario live broadcast, which relates to the field of intelligent routing technology, including collecting live broadcast data, constructing feature vectors, training scene classification models; using time series prediction algorithms to analyze bandwidth requirements and formulate differentiated bandwidth allocation strategies; establishing an edge node load monitoring mechanism, constructing a multi-objective routing optimization model, and realizing real-time scheduling optimization of visitor traffic. This method can effectively improve the bandwidth utilization rate and user experience quality of the live broadcast platform.

Description

Technical Field

[0001] The present invention relates to intelligent routing technology, and in particular to a method and system for bandwidth control and visitor traffic routing optimization for multi-scenario live broadcasting. Background Art

[0002] Traditional bandwidth allocation methods do not take into account the differences in live broadcast scenarios. A unified allocation strategy is difficult to meet the strict requirements of low latency and stability for highly interactive game live broadcasts, which can easily lead to a deterioration in user experience in key scenarios. At the same time, for scenarios with relatively low real-time requirements such as ordinary entertainment live broadcasts and educational live broadcasts, fixed bandwidth allocation methods also cause a waste of resources.

[0003] Secondly, the existing edge node load balancing mechanism often only performs traffic scheduling based on the simple principle of geographical proximity, and fails to fully consider the actual load status and service capabilities of the nodes. This extensive scheduling method easily leads to some nodes being overloaded while other nodes are idle, reducing the overall resource utilization efficiency.

[0004] Thirdly, the system lacks the ability to conduct in-depth analysis and accurate identification of live broadcast scene characteristics, and is unable to perform intelligent resource scheduling based on factors such as network quality requirements and user interaction characteristics in different scenarios. This makes it difficult for the system to make reasonable trade-offs when bandwidth resources are limited, and it is also unable to provide differentiated service guarantees for different types of live broadcast content. Summary of the invention

[0005] The embodiments of the present invention provide a method and system for bandwidth control and visitor traffic routing optimization for multi-scenario live broadcast, which can solve the problems in the prior art.

[0006] According to a first aspect of the embodiments of the present invention,

[0007] Provide bandwidth control and visitor traffic routing optimization methods for multi-scenario live broadcasts, including:

[0008] Collect live broadcast data of multiple live broadcast scenes, the live broadcast data includes network quality parameters, user interaction data and live broadcast content features, wherein the network quality parameters include end-to-end delay, packet loss rate, and jitter, the user interaction data includes viewing time, interaction frequency, and barrage density, and the live broadcast content features include video encoding parameters, audio bit rate, and picture complexity; construct a multidimensional feature vector based on the live broadcast data, use a deep learning algorithm to train a scene classification model, and divide the live broadcast scenes into three types: high-interaction game live broadcast, ordinary entertainment live broadcast, and education live broadcast;

[0009] Use the time series prediction algorithm to analyze the bandwidth demand change trends of various live broadcast scenarios and establish a bandwidth demand prediction model. According to the output results of the bandwidth demand prediction model and the total bandwidth resources of the live broadcast platform, formulate differentiated bandwidth allocation strategies to reserve sufficient bandwidth resources for high-interaction game live broadcasts and allocate dynamically adjusted bandwidth quotas for ordinary entertainment live broadcasts and educational live broadcasts. When fluctuations in bandwidth resources are detected, prioritize the bandwidth requirements of high-interaction game live broadcasts.

[0010] An edge node load status monitoring mechanism is established to collect computing resource occupancy, network throughput, and service quality indicators of each edge node in real time; based on the output results of the scenario classification model and the bandwidth allocation strategy, a multi-objective routing optimization model is constructed to dynamically allocate user access requests to the optimal edge node; the multi-objective routing optimization model comprehensively considers the three goals of node load balance, user access delay, and bandwidth resource utilization, and uses a heuristic algorithm to solve the optimal routing solution to achieve real-time scheduling optimization of visitor traffic.

[0011] The bandwidth demand prediction model is established by using the time series prediction algorithm to analyze the bandwidth demand change trend of various live broadcast scenarios, including:

[0012] The sliding average method is used to extract trend items from the original bandwidth data corresponding to various live broadcast scenarios to obtain trend item data of bandwidth demand; the Fourier transform is used to extract seasonal items from the original bandwidth data, and the frequency components with energy accounting for more than 90% are retained to obtain seasonal item data of bandwidth demand; the trend item data and the seasonal item data are subtracted from the original bandwidth data to obtain random item data of bandwidth demand;

[0013] The trend item data is input into a double-layer long short-term memory network for trend prediction, wherein the number of nodes in the first layer of the double-layer long short-term memory network is 128 and a first dropout layer with a dropout rate of 0.2 is configured, the number of nodes in the second layer of the double-layer long short-term memory network is 64 and a second dropout layer with a dropout rate of 0.2 is configured, and the double-layer long short-term memory network assigns dynamic weights to input features of historical bandwidth data, number of online users, and live content type through an attention mechanism, and outputs trend item prediction data;

[0014] Perform periodic pattern matching on the seasonal item data, calculate the similarity between the time period to be predicted and the historical bandwidth data in the pre-built periodic pattern library based on the dynamic time warping algorithm, select the most similar N periodic data for dynamic adjustment, and output the seasonal item prediction data;

[0015] Performing discrete wavelet transform on the random item data to obtain a high-frequency sub-signal and a low-frequency sub-signal, using a conditional autoregression model to predict the high-frequency sub-signal to obtain a high-frequency prediction result, using a support vector regression model to predict the low-frequency sub-signal to obtain a low-frequency prediction result, and performing inverse wavelet transform on the high-frequency prediction result and the low-frequency prediction result to obtain random item prediction data;

[0016] A bandwidth demand prediction model is established based on the trend item prediction data, the season item prediction data, and the random item prediction data.

[0017] Based on the output of the bandwidth demand prediction model and the total bandwidth resources of the live broadcast platform, differentiated bandwidth allocation strategies are formulated, including:

[0018] Constructing a state vector according to the output result of the bandwidth demand prediction model; constructing an action space for bandwidth adjustment according to the state vector, wherein the action space includes seven discrete decisions: increasing bandwidth by 15%, increasing bandwidth by 10%, increasing bandwidth by 5%, maintaining current bandwidth, decreasing bandwidth by 5%, decreasing bandwidth by 10%, and decreasing bandwidth by 15%;

[0019] Obtain video quality score data, video quality historical score data, buffer ratio data, and delay normalization data of the live broadcast room, and construct an instant reward function according to the difference between the video quality score data and the video quality historical score data, the supplement value of the buffer ratio data, and the supplement value of the delay normalization data;

[0020] Acquire bandwidth utilization data, bandwidth utilization target threshold data, actual revenue data, and bandwidth cost data, and construct a long-term reward function according to the difference between the bandwidth utilization data and the bandwidth utilization target threshold data, and the difference between the actual revenue data and the bandwidth cost data;

[0021] According to the state vector, the action space, the immediate reward function, and the long-term reward function, an independent deep Q network is configured for each live broadcast room, wherein the deep Q network comprises four fully connected layers, the number of nodes in the four fully connected layers is 256, 128, 64, and 32 respectively, and the activation function of the deep Q network is the LeakyReLU function;

[0022] Acquire time series differential error data, calculate sample priority based on the time series differential error data, and store the sample priority data in a playback buffer, wherein the playback buffer has a capacity of 10,000 and a sampling batch size of 64;

[0023] An attention mechanism is constructed based on the sample data in the playback buffer, and the attention mechanism includes a query matrix, a key matrix, and a value matrix. The query matrix is ​​multiplied by the transposed matrix of the key matrix and divided by the square root of the dimension of the key matrix. The calculation result is input into the softmax function to obtain the attention weight, and the attention weight is multiplied by the value matrix to obtain the output result of multi-agent collaborative decision-making, so as to realize bandwidth allocation collaboration among multiple live broadcast rooms.

[0024] Based on the output results of the scenario classification model and the bandwidth allocation strategy, a multi-objective routing optimization model is constructed to dynamically allocate user access requests to the optimal edge node. The method also includes:

[0025] A delay threshold of 50ms is set for the high-interaction game live broadcast, a freeze rate threshold of 0.1% is set for the education live broadcast, and a bit rate threshold of 8Mbps is set for the ordinary entertainment live broadcast, to build a service quality requirement matrix;

[0026] Collect CPU utilization data, memory occupancy data, network bandwidth utilization data, and service quality index data of edge nodes, use a 5-minute sliding time window to calculate the stability scores of the CPU utilization data, the memory occupancy data, the network bandwidth utilization data, and the service quality index data, and obtain the comprehensive health score of the edge node through principal component analysis;

[0027] Obtain user geographic location information data and network measurement data, calculate physical distance data, network hop data, and historical delay data from the user to the edge node based on the user geographic location information data and the network measurement data, use a particle swarm optimization algorithm to adaptively adjust weight coefficients of the physical distance data, the network hop data, and the historical delay data, and construct a user-edge node affinity matrix.

[0028] Based on the output results of the scenario classification model and the bandwidth allocation strategy, a multi-objective routing optimization model is constructed to dynamically allocate user access requests to the optimal edge nodes, including:

[0029] A multi-objective routing optimization model is constructed based on the service quality requirement matrix, the comprehensive health score, and the user-edge node affinity matrix. The optimization objectives of the multi-objective routing optimization model include minimizing user access delay, maximizing edge node resource utilization, and minimizing cross-regional traffic costs. The constraints of the multi-objective routing optimization model include computing resource capacity constraints of edge nodes, network bandwidth capacity constraints, and service quality requirement constraints.

[0030] An improved genetic algorithm is used to solve the multi-objective routing optimization model, the target edge node assigned by the user request is used as the gene position to construct the chromosome by integer coding, the chromosome is selected by the tournament selection method, the single point crossover method is used for crossover operation, and the uniform mutation method is used for mutation operation, wherein the improved genetic algorithm introduces a simulated annealing mechanism on the basis of the genetic algorithm to accept suboptimal solutions;

[0031] A scheduling rule base is established based on the solution results of the multi-objective routing optimization model. The scheduling rule base contains scheduling rules for load fluctuations in different live broadcast scenarios. When a sudden load change is detected, an emergency scheduling mechanism is triggered to prioritize the service quality of high-priority live broadcast scenarios and dynamically allocate user access requests to the optimal edge node.

[0032] The method further includes a user requesting migration:

[0033] The preheating resource scale of the target node is estimated based on historical load data, and the preheating time is dynamically adjusted through an exponential backoff algorithm. Computing resources are pre-allocated, cache is pre-heated, and network links are pre-established at the target node. The source node refers to the edge server node that is currently providing services to users, and the target node refers to the new edge server node selected to undertake service migration;

[0034] Establishing a bidirectional data transmission channel between the source node and the target node, the bidirectional data transmission channel comprising an uplink channel for transmitting control information and a downlink channel for transmitting service data, adopting a three-way handshake mechanism to establish the bidirectional data transmission channel, monitoring the channel status of the bidirectional data transmission channel in real time, and switching to a standby channel when an abnormality is detected based on the channel status;

[0035] Dynamic DNS resolution technology is used to direct newly generated user requests to the target node, a redirection ratio curve is constructed according to an S-shaped function, the number of requests directed to the target node is gradually increased according to the redirection ratio curve, the load state of the target node is monitored, and when the load state is abnormal, the user request is returned to the source node;

[0036] The legacy request of the source node is managed for its life cycle, a maximum waiting time is set, the load status of the source node is monitored, and when the processing time of the legacy request exceeds the maximum waiting time, the legacy request is forcibly migrated to the target node through a compensation mechanism.

[0037] According to a second aspect of the embodiments of the present invention,

[0038] Provide bandwidth control and visitor traffic routing optimization system for multi-scenario live broadcast, including:

[0039] The first unit is used to collect live broadcast data of multiple live broadcast scenes, wherein the live broadcast data includes network quality parameters, user interaction data and live broadcast content features, wherein the network quality parameters include end-to-end delay, packet loss rate and jitter, the user interaction data includes viewing time, interaction frequency and barrage density, and the live broadcast content features include video encoding parameters, audio bit rate and picture complexity; construct a multi-dimensional feature vector based on the live broadcast data, use a deep learning algorithm to train a scene classification model, and divide the live broadcast scenes into three types: high-interaction game live broadcast, ordinary entertainment live broadcast and education live broadcast;

[0040] The second unit is used to use the time series prediction algorithm to analyze the bandwidth demand change trend of various live broadcast scenarios and establish a bandwidth demand prediction model; according to the output results of the bandwidth demand prediction model and the total bandwidth resources of the live broadcast platform, a differentiated bandwidth allocation strategy is formulated to reserve sufficient bandwidth resources for high-interaction game live broadcasts and dynamically adjust bandwidth quotas for ordinary entertainment live broadcasts and educational live broadcasts; when fluctuations in bandwidth resources are detected, the bandwidth demand of high-interaction game live broadcasts is prioritized;

[0041] The third unit is used to establish an edge node load status monitoring mechanism to collect the computing resource occupancy rate, network throughput, and service quality indicators of each edge node in real time; based on the output results of the scenario classification model and the bandwidth allocation strategy, a multi-objective routing optimization model is constructed to dynamically allocate user access requests to the optimal edge node; the multi-objective routing optimization model comprehensively considers the three goals of node load balance, user access delay, and bandwidth resource utilization, and uses a heuristic algorithm to solve the optimal routing solution to achieve real-time scheduling optimization of visitor traffic.

[0042] According to a third aspect of the embodiments of the present invention,

[0043] An electronic device is provided, comprising:

[0044] processor;

[0045] a memory for storing processor-executable instructions;

[0046] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0047] According to a fourth aspect of the embodiments of the present invention,

[0048] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0049] The beneficial effects of this application are as follows:

[0050] 1. By collecting multi-dimensional live broadcast data and building a scene classification model, we can accurately identify and classify different types of live broadcast scenes, providing a reliable decision-making basis for subsequent differentiated bandwidth regulation. The model can accurately distinguish the characteristics of highly interactive game live broadcasts, ordinary entertainment live broadcasts, and educational live broadcasts, effectively improving the pertinence and rationality of bandwidth resource allocation.

[0051] 2. A bandwidth demand prediction model is established based on the time series prediction algorithm, and a differentiated bandwidth allocation strategy is formulated to achieve intelligent regulation and optimal allocation of bandwidth resources. This solution prioritizes bandwidth requirements for highly interactive game live broadcasts, while dynamically adjusting bandwidth for other types of live broadcasts, which not only ensures the service quality of key business scenarios, but also improves the utilization efficiency of bandwidth resources.

[0052] 3. By establishing an edge node load monitoring mechanism and building a multi-objective routing optimization model, intelligent scheduling and optimal allocation of visitor traffic is achieved. The model comprehensively considers multiple objectives such as node load balancing, access latency, and bandwidth utilization, and uses a heuristic algorithm for optimization and solution, effectively improving the overall service quality and user experience of the live broadcast system. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 A flow chart of a method for bandwidth control and visitor traffic routing optimization for multi-scenario live broadcasting according to an embodiment of the present invention;

[0054] Figure 2 A schematic diagram of the structure of a bandwidth control and visitor traffic routing optimization system for multi-scenario live broadcasting according to an embodiment of the present invention. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0056] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0057] Figure 1 Schematic diagram of the process of bandwidth control and visitor traffic routing optimization method for multi-scenario live broadcasting according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0058] S101. Collect live broadcast data of multiple live broadcast scenes, wherein the live broadcast data includes network quality parameters, user interaction data and live broadcast content features, wherein the network quality parameters include end-to-end delay, packet loss rate and jitter, the user interaction data includes viewing time, interaction frequency and barrage density, and the live broadcast content features include video encoding parameters, audio bit rate and picture complexity; construct a multi-dimensional feature vector based on the live broadcast data, use a deep learning algorithm to train a scene classification model, and divide the live broadcast scenes into three types: high-interaction game live broadcast, ordinary entertainment live broadcast and education live broadcast;

[0059] S102. Use the time series prediction algorithm to analyze the bandwidth demand change trend of various live broadcast scenarios and establish a bandwidth demand prediction model; according to the output results of the bandwidth demand prediction model and the total bandwidth resources of the live broadcast platform, formulate a differentiated bandwidth allocation strategy to reserve sufficient bandwidth resources for high-interaction game live broadcasts and allocate dynamically adjusted bandwidth quotas for ordinary entertainment live broadcasts and educational live broadcasts; when fluctuations in bandwidth resources are detected, give priority to ensuring the bandwidth demand of high-interaction game live broadcasts;

[0060] S103. Establish an edge node load status monitoring mechanism to collect the computing resource occupancy rate, network throughput, and service quality indicators of each edge node in real time; based on the output results of the scenario classification model and the bandwidth allocation strategy, construct a multi-objective routing optimization model to dynamically allocate user access requests to the optimal edge node; the multi-objective routing optimization model comprehensively considers the three goals of node load balance, user access delay, and bandwidth resource utilization, and uses a heuristic algorithm to solve the optimal routing solution to achieve real-time scheduling optimization of visitor traffic.

[0061] This embodiment provides a bandwidth control and visitor traffic routing optimization method for multi-scenario live broadcast. The method first collects live broadcast data of multiple live broadcast scenarios, including network quality parameters, user interaction data, and live broadcast content characteristics. Network quality parameters include end-to-end delay, packet loss rate, and jitter. User interaction data includes viewing time, interaction frequency, and bullet screen density. Live broadcast content characteristics include video encoding parameters, audio bit rate, and picture complexity.

[0062] Based on the collected live broadcast data, a multi-dimensional feature vector is constructed. The feature vector contains the above parameters, and each parameter corresponds to a dimension. For example, a 9-dimensional feature vector can be constructed, corresponding to the nine parameters of end-to-end delay, packet loss rate, jitter, viewing time, interaction frequency, bullet screen density, video encoding parameters, audio bit rate, and picture complexity.

[0063] Then, we use the deep learning algorithm to train the scene classification model and divide the live broadcast scenes into three types: highly interactive game live broadcast, ordinary entertainment live broadcast, and educational live broadcast. Specifically, we can use the convolutional neural network (CNN) model, with the input as the above 9-dimensional feature vector and the output as the probability distribution of the three scene types. When training the model, we can use a large amount of live broadcast data with labeled scene types for supervised learning.

[0064] Next, we use the time series prediction algorithm to analyze the bandwidth demand change trends of various live broadcast scenarios and establish a bandwidth demand prediction model. We can use the long short-term memory network (LSTM) model, with the input being the bandwidth demand sequence data of a period of history (such as the past 7 days) and the output being the bandwidth demand forecast value for a period of time in the future (such as the next 24 hours). We train the LSTM model for each of the three scenario types.

[0065] According to the output results of the bandwidth demand prediction model and the total bandwidth resources of the live broadcast platform, a differentiated bandwidth allocation strategy is formulated. For example, assuming that the total bandwidth resources are 100Gbps, the bandwidth requirements for high-interaction game live broadcast, ordinary entertainment live broadcast, and education live broadcast are predicted to be 50Gbps, 30Gbps, and 20Gbps, respectively. Then, 55Gbps of bandwidth can be reserved for high-interaction game live broadcast, 25Gbps of bandwidth can be allocated for ordinary entertainment live broadcast, and 20Gbps of bandwidth can be allocated for education live broadcast. When fluctuations in bandwidth resources are detected, the bandwidth requirements of high-interaction game live broadcast are prioritized, and bandwidth resources can be drawn from the other two types of live broadcasts when necessary.

[0066] Establish an edge node load status monitoring mechanism to collect computing resource utilization, network throughput, and service quality indicators of each edge node in real time. A monitoring agent can be deployed on each edge node to collect CPU usage, memory usage, disk I / O, network throughput, service response time and other indicators regularly (e.g., every 5 minutes) and report them to the central control system.

[0067] Based on the output results of the scenario classification model and the bandwidth allocation strategy, a multi-objective routing optimization model is constructed to dynamically allocate user access requests to the optimal edge node. The multi-objective routing optimization model comprehensively considers the three goals of node load balance, user access delay, and bandwidth resource utilization. Genetic algorithms can be used to solve the optimal routing solution.

[0068] The specific implementation process is as follows:

[0069] First, we define the chromosome encoding method. Assuming there are N user requests and M edge nodes, the length of the chromosome is N, and the value range of each gene is from 1 to M, indicating which edge node the user request is assigned to.

[0070] Then, define the fitness function. The fitness function consists of three parts: node load balance, user access delay, and bandwidth resource utilization. Node load balance can be measured by variance. The smaller the variance, the more balanced the load. User access delay can calculate the network distance based on the user's geographic location and the edge node location. Bandwidth resource utilization can calculate the ratio of each node's allocated bandwidth to the actual bandwidth used. The three objectives can be considered comprehensively by weighted summation.

[0071] Next, perform genetic operations such as population initialization, selection, crossover, and mutation. The population size can be set to 100 and the number of generations can be set to 1000. In each generation, select individuals with higher fitness for crossover and mutation to produce the next generation of population. Finally, select the individual with the highest fitness as the optimal solution.

[0072] Finally, according to the optimal solution, user requests are assigned to the corresponding edge nodes to achieve real-time scheduling optimization of visitor traffic. The system can re-execute the optimization process at regular intervals (such as 5 minutes) to adapt to the dynamic changes of the network environment.

[0073] The beneficial effects of this method include:

[0074] 1. Through scene classification and differentiated bandwidth allocation, the efficiency of bandwidth resource utilization is improved, and the service quality of highly interactive live broadcast scenes is guaranteed. 2. Multi-objective routing optimization is adopted to achieve a comprehensive balance of edge node load balancing, user access delay minimization and efficient utilization of bandwidth resources. 3. The dynamic adjustment strategy can quickly respond to changes in the network environment, improving the robustness and reliability of the system.

[0075] In an optional implementation, using a time series prediction algorithm to analyze the bandwidth demand change trend of various live broadcast scenarios, and establishing a bandwidth demand prediction model includes:

[0076] The sliding average method is used to extract trend items from the original bandwidth data corresponding to various live broadcast scenarios to obtain trend item data of bandwidth demand; the Fourier transform is used to extract seasonal items from the original bandwidth data, and the frequency components with energy accounting for more than 90% are retained to obtain seasonal item data of bandwidth demand; the trend item data and the seasonal item data are subtracted from the original bandwidth data to obtain random item data of bandwidth demand;

[0077] The trend item data is input into a double-layer long short-term memory network for trend prediction, wherein the number of nodes in the first layer of the double-layer long short-term memory network is 128 and a first dropout layer with a dropout rate of 0.2 is configured, the number of nodes in the second layer of the double-layer long short-term memory network is 64 and a second dropout layer with a dropout rate of 0.2 is configured, and the double-layer long short-term memory network assigns dynamic weights to input features of historical bandwidth data, number of online users, and live content type through an attention mechanism, and outputs trend item prediction data;

[0078] Perform periodic pattern matching on the seasonal item data, calculate the similarity between the time period to be predicted and the historical bandwidth data in the pre-built periodic pattern library based on the dynamic time warping algorithm, select the most similar N periodic data for dynamic adjustment, and output the seasonal item prediction data;

[0079] Performing discrete wavelet transform on the random item data to obtain a high-frequency sub-signal and a low-frequency sub-signal, using a conditional autoregression model to predict the high-frequency sub-signal to obtain a high-frequency prediction result, using a support vector regression model to predict the low-frequency sub-signal to obtain a low-frequency prediction result, and performing inverse wavelet transform on the high-frequency prediction result and the low-frequency prediction result to obtain random item prediction data;

[0080] A bandwidth demand prediction model is established based on the trend item prediction data, the season item prediction data, and the random item prediction data.

[0081] In the live broadcast scenario, by predicting and analyzing the bandwidth demand, we first need to collect historical bandwidth data in different types of live broadcast scenarios, including the number of online viewers, live broadcast content type and other related information. Based on the collected original bandwidth data, we extract three components: trend item, seasonal item and random item.

[0082] For trend item extraction, a sliding window with a 5-point moving average is used to smooth the original bandwidth data. The window slides forward one time point each time to obtain the trend change of bandwidth demand. For example, in a game live broadcast scenario, the original bandwidth data is [2.1, 2.3, 2.0, 2.4, 2.2, 2.5, 2.3] Mbps. After sliding average processing, the trend item data is [2.2, 2.28, 2.28, 2.34] Mbps.

[0083] For seasonal term extraction, the original bandwidth data is subjected to fast Fourier transform to analyze its spectral characteristics. By calculating the energy proportion of each frequency component, the main frequency components with cumulative energy exceeding 90% are retained, such as retaining significant periodic features such as daily cycle (24 hours) and weekly cycle (7 days), removing high-frequency noise, and obtaining the seasonal variation pattern of bandwidth demand.

[0084] The random item data is obtained by subtracting the trend item and the seasonal item from the original data. The random item data is decomposed by three layers of discrete wavelet to obtain high-frequency sub-signals and low-frequency sub-signals. The conditional autoregressive model is used to model the high-frequency signal, and the historical data with a lag of 3 periods is input for prediction; the support vector regression model of the radial basis kernel function is used to model the low-frequency signal, and the parameter configuration with a penalty factor of 1.0 and an epsilon of 0.1 is selected. Finally, the high-frequency and low-frequency prediction results are reconstructed by wavelet to obtain the random item prediction value.

[0085] For trend prediction, a two-layer long short-term memory network is constructed, with 128 nodes in the first layer and a dropout of 0.2, and 64 nodes in the second layer and a dropout of 0.2. The network input includes bandwidth data, number of online users, live broadcast type and other features of the past 24 hours, and the dynamic weight of each feature is calculated through the attention layer. For example, the weight of the number of online users is 0.5, the weight of bandwidth data is 0.3, and the weight of live broadcast type is 0.2, and the weighted network input is obtained.

[0086] For seasonal forecasting, a periodic pattern library is pre-built to store typical periodic patterns of historical data. The dynamic time warping algorithm is used to calculate the similarity between the period to be predicted and the pattern library, and the three most similar periodic data are selected. The selected periodic data are dynamically adjusted according to the actual situation of the current period to obtain the seasonal item forecast value.

[0087] Finally, the trend forecast value, seasonal forecast value and random forecast value are superimposed to obtain the overall forecast result of bandwidth demand. Through practical application verification, the average forecast error of the forecast model is within 10%.

[0088] The beneficial effects of this method include: 1) through the decomposition-prediction-combination framework, the trend changes, periodic characteristics and random fluctuations of bandwidth demand are effectively captured, thereby improving the prediction accuracy; 2) by combining deep learning with a variety of machine learning algorithms, multi-dimensional input features are fully utilized to enhance the generalization ability of the model; 3) the prediction results can provide a decision-making basis for network resource scheduling, helping to achieve accurate allocation of bandwidth resources and cost optimization.

[0089] In an optional implementation, based on the output result of the bandwidth demand prediction model and combined with the total bandwidth resources of the live broadcast platform, a differentiated bandwidth allocation strategy is formulated, including:

[0090] Constructing a state vector according to the output result of the bandwidth demand prediction model; constructing an action space for bandwidth adjustment according to the state vector, wherein the action space includes seven discrete decisions: increasing bandwidth by 15%, increasing bandwidth by 10%, increasing bandwidth by 5%, maintaining current bandwidth, decreasing bandwidth by 5%, decreasing bandwidth by 10%, and decreasing bandwidth by 15%;

[0091] Obtain video quality score data, video quality historical score data, buffer ratio data, and delay normalization data of the live broadcast room, and construct an instant reward function according to the difference between the video quality score data and the video quality historical score data, the supplement value of the buffer ratio data, and the supplement value of the delay normalization data;

[0092] Acquire bandwidth utilization data, bandwidth utilization target threshold data, actual revenue data, and bandwidth cost data, and construct a long-term reward function according to the difference between the bandwidth utilization data and the bandwidth utilization target threshold data, and the difference between the actual revenue data and the bandwidth cost data;

[0093] According to the state vector, the action space, the immediate reward function, and the long-term reward function, an independent deep Q network is configured for each live broadcast room, wherein the deep Q network comprises four fully connected layers, the number of nodes in the four fully connected layers is 256, 128, 64, and 32 respectively, and the activation function of the deep Q network is the LeakyReLU function;

[0094] Acquire time series differential error data, calculate sample priority based on the time series differential error data, and store the sample priority data in a playback buffer, wherein the playback buffer has a capacity of 10,000 and a sampling batch size of 64;

[0095] An attention mechanism is constructed based on the sample data in the playback buffer, and the attention mechanism includes a query matrix, a key matrix, and a value matrix. The query matrix is ​​multiplied by the transposed matrix of the key matrix and divided by the square root of the dimension of the key matrix. The calculation result is input into the softmax function to obtain the attention weight, and the attention weight is multiplied by the value matrix to obtain the output result of multi-agent collaborative decision-making, so as to realize bandwidth allocation collaboration among multiple live broadcast rooms.

[0096] First, obtain the real-time data of each live broadcast room in the live broadcast platform, including current bandwidth usage, number of viewers, video bit rate and other information. Based on these data, construct the input feature vector of the bandwidth demand prediction model. The feature vector contains multi-dimensional features such as the current bandwidth usage of the live broadcast room, the mean and variance of the bandwidth usage in the past hour, the number of viewers, and the video bit rate. Input the feature vector into the pre-trained bandwidth demand prediction model to obtain the bandwidth demand prediction value of the live broadcast room in the future.

[0097] The predicted bandwidth demand value of each live broadcast room is normalized to construct a state vector. The state vector contains information such as the current bandwidth occupancy ratio, predicted bandwidth demand ratio, and bandwidth utilization of the live broadcast room. For example, if the current bandwidth occupancy of a live broadcast room is 100Mbps and the predicted bandwidth demand is 120Mbps, the normalized state vector components are 0.5 and 0.6 respectively.

[0098] The action space of bandwidth adjustment is constructed based on the state vector, which contains seven discrete decision options: increase bandwidth by 15%, 10%, and 5%, keep the current bandwidth unchanged, and reduce bandwidth by 5%, 10%, and 15%. Each action corresponds to an independent Q value, which is used to evaluate the long-term benefits of the action.

[0099] Construct an immediate reward function, the input includes: the difference between the video quality score and the historical average (such as the current score of 8.5, the historical average score of 8, and the difference of 0.5), the buffer ratio compensation value (such as the buffer ratio of 0.2, the compensation value is 0.8), and the compensation value after delay normalization (such as the standardized delay of 0.3, the compensation value is 0.7). These indicators reflect the direct impact of the current bandwidth allocation on the user experience.

[0100] The long-term reward function is constructed with inputs including: the difference between bandwidth utilization and target threshold (e.g. current utilization 0.75, target threshold 0.8, difference is -0.05), the difference between actual revenue and bandwidth cost (e.g. revenue 1,000 yuan, cost 800 yuan, difference is 200 yuan). These indicators reflect the long-term economic benefits of bandwidth allocation.

[0101] An independent deep Q network is configured for each live broadcast room, which contains 4 fully connected layers with 256, 128, 64, and 32 nodes respectively. LeakyReLU is used as the activation function to avoid the problem of neuron death. The network input is the state vector and the output is the Q value estimate of each action.

[0102] Calculate the temporal difference error to evaluate the accuracy of the prediction. For example, if the predicted Q value of a state-action pair is 10 and the actual benefit is 12, the temporal difference error is 2. Calculate the sample priority based on the error size. The higher the priority, the greater the probability of being selected during training. Store the samples and their priorities in a playback buffer with a capacity of 10,000, and randomly sample 64 samples each time for training.

[0103] Construct an attention mechanism to achieve multi-agent collaboration. The dimensions of the query matrix, key matrix, and value matrix match the dimensions of the state vector. The attention weights are obtained through matrix operations to reflect the correlation between different live broadcast rooms. The decisions of each live broadcast room are weighted and integrated according to the attention weights to achieve overall optimization of bandwidth allocation.

[0104] Beneficial effects include:

[0105] 1. Through deep reinforcement learning and attention mechanism, the bandwidth allocation of live broadcast room is intelligent and automated, which improves the efficiency of resource utilization. 2. By combining immediate rewards and long-term rewards, the economic benefits are maximized while ensuring user experience. 3. The introduction of sample priority mechanism and multi-agent collaborative decision-making improves the learning efficiency of the model and the accuracy of decision-making.

[0106] In an optional implementation, based on the output results of the scenario classification model and the bandwidth allocation strategy, a multi-objective routing optimization model is constructed, and before the user access request is dynamically allocated to the optimal edge node, the method further includes:

[0107] A delay threshold of 50ms is set for the high-interaction game live broadcast, a freeze rate threshold of 0.1% is set for the education live broadcast, and a bit rate threshold of 8Mbps is set for the ordinary entertainment live broadcast, to build a service quality requirement matrix;

[0108] Collect CPU utilization data, memory occupancy data, network bandwidth utilization data, and service quality index data of edge nodes, use a 5-minute sliding time window to calculate the stability scores of the CPU utilization data, the memory occupancy data, the network bandwidth utilization data, and the service quality index data, and obtain the comprehensive health score of the edge node through principal component analysis;

[0109] Obtain user geographic location information data and network measurement data, calculate physical distance data, network hop data, and historical delay data from the user to the edge node based on the user geographic location information data and the network measurement data, use a particle swarm optimization algorithm to adaptively adjust weight coefficients of the physical distance data, the network hop data, and the historical delay data, and construct a user-edge node affinity matrix.

[0110] In actual applications, we first set differentiated service quality thresholds for different types of live broadcast scenarios. For highly interactive game live broadcasts, due to their high sensitivity to delay, the end-to-end delay threshold is set to 50 milliseconds; for educational live broadcasts, to ensure teaching effectiveness, the freeze rate threshold is set to 0.1%; and for ordinary entertainment live broadcasts, the main focus is on picture quality, so the bit rate threshold is set to 8Mbps. Based on these threshold requirements, a service quality requirement matrix is ​​constructed for subsequent resource scheduling decisions.

[0111] In terms of edge node health assessment, the system collects performance indicator data of edge nodes every 5 minutes, including CPU utilization, memory occupancy, network bandwidth utilization, and service quality indicators. Taking CPU utilization as an example, a 5-minute sliding time window is used to calculate the mean, variance and other statistical characteristics of the data in the window, and the stability score of the CPU indicator is obtained comprehensively. The stability scores of other indicators are calculated in the same way. For example, the mean CPU utilization of an edge node in a time window is 65%, and the variance is 3%, which can obtain a higher stability score of 0.85. The stability scores of all performance indicators are used as input, and the main features are extracted through the principal component analysis method, and finally the comprehensive health score of the edge node is obtained.

[0112] For the affinity evaluation between users and edge nodes, we first obtain the user's geographic location information (such as longitude and latitude coordinates) and network measurement data. Based on the user's location information, calculate the physical distance from the user to each edge node. For example, if the user is located at 39.9°N, 116.3°E, and an edge node is located at 39.8°N, 116.4°E, the actual physical distance can be calculated. At the same time, the network hop count and historical delay data are obtained through network measurement. For example, if the network hop count is 3 hops, the historical average delay is 20ms. The particle swarm optimization algorithm is used to dynamically adjust the weight coefficients of the three indicators of physical distance, network hop count, and historical delay according to the actual network conditions, and finally construct the user-edge node affinity matrix.

[0113] This technical solution has the following beneficial effects:

[0114] 1. By setting differentiated service quality thresholds, the requirements of different types of live broadcast services are met and the user experience is improved; the edge node health assessment mechanism can detect performance anomalies in a timely manner and ensure service stability. 2. The sliding time window and principal component analysis method are used to comprehensively evaluate the operating status of edge nodes; the comprehensive score based on multi-dimensional indicators is more objective and accurate, avoiding the one-sidedness caused by a single indicator. 3. The particle swarm optimization algorithm is introduced to dynamically adjust the weight coefficient, so that the user-edge node affinity assessment is more in line with the actual network environment; based on the comprehensive consideration of physical distance, network hop count, and historical delay, better user request scheduling can be achieved.

[0115] In an optional implementation, based on the output results of the scenario classification model and the bandwidth allocation strategy, a multi-objective routing optimization model is constructed to dynamically allocate user access requests to the optimal edge node, including:

[0116] A multi-objective routing optimization model is constructed based on the service quality requirement matrix, the comprehensive health score, and the user-edge node affinity matrix. The optimization objectives of the multi-objective routing optimization model include minimizing user access delay, maximizing edge node resource utilization, and minimizing cross-regional traffic costs. The constraints of the multi-objective routing optimization model include computing resource capacity constraints of edge nodes, network bandwidth capacity constraints, and service quality requirement constraints.

[0117] An improved genetic algorithm is used to solve the multi-objective routing optimization model, the target edge node assigned by the user request is used as the gene position to construct the chromosome by integer coding, the chromosome is selected by the tournament selection method, the single point crossover method is used for crossover operation, and the uniform mutation method is used for mutation operation, wherein the improved genetic algorithm introduces a simulated annealing mechanism on the basis of the genetic algorithm to accept suboptimal solutions;

[0118] A scheduling rule base is established based on the solution results of the multi-objective routing optimization model. The scheduling rule base contains scheduling rules for load fluctuations in different live broadcast scenarios. When a sudden load change is detected, an emergency scheduling mechanism is triggered to prioritize the service quality of high-priority live broadcast scenarios and dynamically allocate user access requests to the optimal edge node.

[0119] In order to achieve the output results and bandwidth allocation strategy based on the scenario classification model, a multi-objective routing optimization model is constructed to dynamically allocate user access requests to the optimal edge node. The following specific implementation methods can be adopted:

[0120] First, a multi-objective routing optimization model is constructed based on the service quality requirement matrix, comprehensive health score, and user-edge node affinity matrix. The service quality requirement matrix contains the requirements of different live broadcast scenarios for indicators such as delay and bandwidth. For example, game live broadcast has higher requirements for delay, which is set to within 50ms; the comprehensive health score reflects the overall status of the edge node, including CPU utilization, memory utilization, network status, etc. For example, the score is 90 points when the CPU utilization is less than 60%, the memory utilization is less than 70%, and the network delay is less than 20ms; the user-edge node affinity matrix represents the network distance between the user and the edge node, for example, the RTT of user A to edge node 1 is 15ms, and the RTT to edge node 2 is 25ms. The optimization goals of the multi-objective routing optimization model include: minimizing user access delay, such as controlling the average access delay within 30ms; maximizing edge node resource utilization, such as increasing the average CPU and memory utilization to 70%; minimizing cross-regional traffic costs, such as controlling the cross-regional traffic ratio within 20%. The constraints include: computing resource capacity constraints of edge nodes, such as a single edge node supporting up to 1,000 concurrent users; network bandwidth capacity constraints, such as the upper limit of the export bandwidth of a single edge node is 10Gbps; service quality requirement constraints, such as the end-to-end delay of game live broadcasts not exceeding 100ms.

[0121] Next, an improved genetic algorithm is used to solve the multi-objective routing optimization model. The target edge node assigned by the user request is used as the gene position to construct the chromosome using integer coding. For example, chromosome [1,2,3,1,2] means that 5 users are assigned to edge nodes 1, 2, 3, 1, and 2 respectively. The chromosome is selected by the tournament selection method, for example, 3 chromosomes are randomly selected from the population each time, and the one with the highest fitness is selected to enter the next generation. The single-point crossover method is used for crossover operation, for example, crossover is performed at a random position of the chromosome to generate a new offspring chromosome. The uniform mutation method is used for mutation operation, for example, each gene in the chromosome is mutated with a probability of 0.1. On the basis of the genetic algorithm, a simulated annealing mechanism is introduced to accept suboptimal solutions, for example, solutions with poor fitness are accepted with a certain probability to avoid falling into the local optimum. Specifically, the initial temperature is set to 100, the termination temperature is set to 0.1, and the cooling coefficient is set to 0.95. In each iteration, if the new solution is worse than the current solution, the new solution is accepted with a probability of exp((f(new)-f(current)) / T), where f() is the fitness function, T is the current temperature, new represents the new solution, and current represents the current solution.

[0122] Then, a scheduling rule base is established based on the solution results of the multi-objective routing optimization model. The scheduling rule base contains scheduling rules for load fluctuations in different live broadcast scenarios. For example, for game live broadcast scenarios, when the CPU utilization of an edge node is detected to exceed 80%, the new user request is allocated to the adjacent edge node with lighter load; for large-scale event live broadcast scenarios, when the number of users is detected to increase by more than 50%, a new edge node is started for capacity expansion. When a sudden load change is detected, the emergency scheduling mechanism is triggered to prioritize the service quality of high-priority live broadcast scenarios. For example, when the bandwidth utilization of an edge node is detected to exceed 90%, low-priority on-demand traffic is first migrated to other nodes to reserve bandwidth resources for high-priority live broadcast traffic. Finally, according to the rules in the scheduling rule base, user access requests are dynamically allocated to the optimal edge node. For example, for newly connected user requests, the three nearest edge nodes are first selected based on the user location information, and then the node with the highest comprehensive score is selected as the optimal edge node for allocation considering the load and network conditions of these nodes.

[0123] This method has the following beneficial effects:

[0124] 1. By constructing a multi-objective routing optimization model and comprehensively considering multiple optimization objectives such as user access delay, edge node resource utilization, and cross-regional traffic cost, it is possible to achieve a good balance between multiple performance indicators and improve the overall system operation efficiency and service quality.

[0125] 2. Using an improved genetic algorithm to solve the multi-objective routing optimization model and introducing a simulated annealing mechanism to accept suboptimal solutions can effectively avoid falling into local optimal solutions, improve the global optimization capability, and obtain a better scheduling solution.

[0126] 3. Establish a scheduling rule library for different live broadcast scenarios and set up an emergency scheduling mechanism to quickly respond to load fluctuations, dynamically adjust resource allocation strategies, ensure the service quality of key businesses, and improve the robustness and reliability of the system.

[0127] In an optional implementation, the method further includes a user requesting migration:

[0128] The preheating resource scale of the target node is estimated based on historical load data, and the preheating time is dynamically adjusted through an exponential backoff algorithm. Computing resources are pre-allocated, cache is pre-heated, and network links are pre-established at the target node. The source node refers to the edge server node that is currently providing services to users, and the target node refers to the new edge server node selected to undertake service migration;

[0129] Establishing a bidirectional data transmission channel between the source node and the target node, the bidirectional data transmission channel comprising an uplink channel for transmitting control information and a downlink channel for transmitting service data, adopting a three-way handshake mechanism to establish the bidirectional data transmission channel, monitoring the channel status of the bidirectional data transmission channel in real time, and switching to a standby channel when an abnormality is detected based on the channel status;

[0130] Dynamic DNS resolution technology is used to direct newly generated user requests to the target node, a redirection ratio curve is constructed according to an S-shaped function, the number of requests directed to the target node is gradually increased according to the redirection ratio curve, the load state of the target node is monitored, and when the load state is abnormal, the user request is returned to the source node;

[0131] The legacy request of the source node is managed for its life cycle, a maximum waiting time is set, the load status of the source node is monitored, and when the processing time of the legacy request exceeds the maximum waiting time, the legacy request is forcibly migrated to the target node through a compensation mechanism.

[0132] During the user request migration process, the resource scale required for the target node is first estimated based on historical load data. By analyzing indicators such as CPU usage, memory usage, and network bandwidth in historical load data, and combining the Gaussian mixture model to model resource usage patterns, the resource scale that the target node needs to preheat is predicted. For example, when historical data shows that the peak CPU usage reaches 80% and the memory usage is 6GB, the preheating resource scale can be set to 4 CPU cores and 8GB of memory.

[0133] The exponential backoff algorithm is used to dynamically adjust the warm-up time. The initial warm-up time is set to 30 seconds. If it is detected that the target node resources are not ready, the warm-up time is doubled to 60 seconds, and a maximum of 3 retries are made. During the warm-up phase, the target node pre-allocates computing resources, such as allocating virtual CPUs and memory in advance; pre-heats the cache to load hot data into the memory in advance; and pre-establishes network links, including configuring load balancing and firewall rules.

[0134] Then, a two-way data transmission channel is established between the source node and the target node. The uplink channel is used to transmit control information such as heartbeat detection and resource status, and the TCP protocol is used to ensure reliable transmission; the downlink channel is used to transmit actual business data, and the UDP protocol can be used to improve transmission efficiency. The channel is established through a three-way handshake: first, the source node sends a SYN packet, the target node replies with a SYN-ACK packet, and the source node finally sends an ACK packet to complete the connection. The channel status is monitored in real time. When the packet loss rate exceeds 1% or the delay exceeds 100ms, it automatically switches to the pre-configured backup channel.

[0135] Use dynamic DNS resolution technology to gradually direct new user requests to the target node. Construct a redirection ratio curve based on the S-shaped function, slowly increase the redirection ratio in the early stage of migration, quickly increase it in the middle stage, and slow down again in the later stage. Specifically, redirect 10% of requests in the first 10 minutes, increase to 50% in 10-20 minutes, increase to 90% in 20-30 minutes, and complete the redirection of the remaining requests in the last 10 minutes. During this period, continuously monitor indicators such as the target node's CPU usage and response time. When the CPU usage exceeds 95% or the response time exceeds 3 seconds, the new request will be rolled back to the source node.

[0136] Perform lifecycle management on the legacy requests on the source node, and set the maximum waiting time to 5 minutes. Continuously monitor the load status of the source node, including indicators such as request queue length and processing delay. When the processing time of the legacy request exceeds the maximum waiting time, start the compensation mechanism: package the request context information and send it to the target node through a reliable transmission channel; the target node restores the request context and continues processing to ensure that the request processing is not interrupted.

[0137] This solution has the following advantages: First, through precise resource preheating and dynamic adjustment mechanisms, the success rate of service migration is significantly improved and the risk of service interruption is reduced; second, the design of the two-way data transmission channel and the backup channel mechanism ensure the reliability and real-time performance of data transmission and improve the stability of the migration process; finally, the progressive migration solution based on the redirection ratio curve, combined with timely anomaly detection and fallback mechanism, achieves smooth service migration and effectively guarantees user experience.

[0138] Figure 2FIG. 1 is a schematic diagram of a bandwidth control and visitor traffic routing optimization system for multi-scenario live broadcasting according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0139] The first unit is used to collect live broadcast data of multiple live broadcast scenes, wherein the live broadcast data includes network quality parameters, user interaction data and live broadcast content features, wherein the network quality parameters include end-to-end delay, packet loss rate and jitter, the user interaction data includes viewing time, interaction frequency and barrage density, and the live broadcast content features include video encoding parameters, audio bit rate and picture complexity; construct a multi-dimensional feature vector based on the live broadcast data, use a deep learning algorithm to train a scene classification model, and divide the live broadcast scenes into three types: high-interaction game live broadcast, ordinary entertainment live broadcast and education live broadcast;

[0140] The second unit is used to use the time series prediction algorithm to analyze the bandwidth demand change trend of various live broadcast scenarios and establish a bandwidth demand prediction model; according to the output results of the bandwidth demand prediction model and the total bandwidth resources of the live broadcast platform, a differentiated bandwidth allocation strategy is formulated to reserve sufficient bandwidth resources for high-interaction game live broadcasts and dynamically adjust bandwidth quotas for ordinary entertainment live broadcasts and educational live broadcasts; when fluctuations in bandwidth resources are detected, the bandwidth demand of high-interaction game live broadcasts is prioritized;

[0141] The third unit is used to establish an edge node load status monitoring mechanism to collect the computing resource occupancy rate, network throughput, and service quality indicators of each edge node in real time; based on the output results of the scenario classification model and the bandwidth allocation strategy, a multi-objective routing optimization model is constructed to dynamically allocate user access requests to the optimal edge node; the multi-objective routing optimization model comprehensively considers the three goals of node load balance, user access delay, and bandwidth resource utilization, and uses a heuristic algorithm to solve the optimal routing solution to achieve real-time scheduling optimization of visitor traffic.

[0142] According to a third aspect of the embodiments of the present invention,

[0143] An electronic device is provided, comprising:

[0144] processor;

[0145] a memory for storing processor-executable instructions;

[0146] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0147] According to a fourth aspect of the embodiments of the present invention,

[0148] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0149] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0150] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for bandwidth control and visitor traffic routing optimization for multi-scenario live broadcast, characterized in that: include: Collecting live broadcast data of multiple live broadcast scenes, wherein the live broadcast data includes network quality parameters, user interaction data, and live broadcast content features, wherein the network quality parameters include end-to-end delay, packet loss rate, and jitter, the user interaction data includes viewing time, interaction frequency, and barrage density, and the live broadcast content features include video encoding parameters, audio bit rate, and picture complexity; Constructing a multi-dimensional feature vector based on the live broadcast data, using a deep learning algorithm to train a scene classification model, and dividing the live broadcast scenes into three types: high-interaction game live broadcast, ordinary entertainment live broadcast, and education live broadcast; Use the time series prediction algorithm to analyze the bandwidth demand change trends of various live broadcast scenarios and establish a bandwidth demand prediction model. According to the output results of the bandwidth demand prediction model and the total bandwidth resources of the live broadcast platform, formulate differentiated bandwidth allocation strategies to reserve sufficient bandwidth resources for high-interaction game live broadcasts and allocate dynamically adjusted bandwidth quotas for ordinary entertainment live broadcasts and educational live broadcasts. When fluctuations in bandwidth resources are detected, prioritize the bandwidth requirements of high-interaction game live broadcasts. Establish an edge node load status monitoring mechanism to collect computing resource occupancy, network throughput, and service quality indicators of each edge node in real time; Based on the output results of the scenario classification model and the bandwidth allocation strategy, a multi-objective routing optimization model is constructed to dynamically allocate user access requests to the optimal edge node; the multi-objective routing optimization model comprehensively considers the three goals of node load balance, user access delay, and bandwidth resource utilization, and uses a heuristic algorithm to solve the optimal routing solution to achieve real-time scheduling optimization of visitor traffic; The bandwidth demand prediction model is established by using the time series prediction algorithm to analyze the bandwidth demand change trend of various live broadcast scenarios, including: The sliding average method is used to extract trend items from the original bandwidth data corresponding to various live broadcast scenarios to obtain trend item data of bandwidth demand; the Fourier transform is used to extract seasonal items from the original bandwidth data, and the frequency components with energy accounting for more than 90% are retained to obtain seasonal item data of bandwidth demand; the trend item data and the seasonal item data are subtracted from the original bandwidth data to obtain random item data of bandwidth demand; The trend item data is input into a double-layer long short-term memory network for trend prediction, wherein the number of nodes in the first layer of the double-layer long short-term memory network is 128 and a first dropout layer with a dropout rate of 0.2 is configured, the number of nodes in the second layer of the double-layer long short-term memory network is 64 and a second dropout layer with a dropout rate of 0.2 is configured, and the double-layer long short-term memory network assigns dynamic weights to input features of historical bandwidth data, number of online users, and live content type through an attention mechanism, and outputs trend item prediction data; Perform periodic pattern matching on the seasonal item data, calculate the similarity between the time period to be predicted and the historical bandwidth data in the pre-built periodic pattern library based on the dynamic time warping algorithm, select the most similar N periodic data for dynamic adjustment, and output the seasonal item prediction data; Performing discrete wavelet transform on the random item data to obtain a high-frequency sub-signal and a low-frequency sub-signal, using a conditional autoregression model to predict the high-frequency sub-signal to obtain a high-frequency prediction result, using a support vector regression model to predict the low-frequency sub-signal to obtain a low-frequency prediction result, and performing inverse wavelet transform on the high-frequency prediction result and the low-frequency prediction result to obtain random item prediction data; A bandwidth demand prediction model is established based on the trend item prediction data, the season item prediction data, and the random item prediction data.

2. The method according to claim 1, characterized in that Based on the output of the bandwidth demand prediction model and the total bandwidth resources of the live broadcast platform, differentiated bandwidth allocation strategies are formulated, including: Constructing a state vector according to the output result of the bandwidth demand prediction model; constructing an action space for bandwidth adjustment according to the state vector, wherein the action space includes seven discrete decisions: increasing bandwidth by 15%, increasing bandwidth by 10%, increasing bandwidth by 5%, maintaining current bandwidth, decreasing bandwidth by 5%, decreasing bandwidth by 10%, and decreasing bandwidth by 15%; Obtain video quality score data, video quality historical score data, buffer ratio data, and delay normalization data of the live broadcast room, and construct an instant reward function according to the difference between the video quality score data and the video quality historical score data, the supplement value of the buffer ratio data, and the supplement value of the delay normalization data; Acquire bandwidth utilization data, bandwidth utilization target threshold data, actual revenue data, and bandwidth cost data, and construct a long-term reward function according to the difference between the bandwidth utilization data and the bandwidth utilization target threshold data, and the difference between the actual revenue data and the bandwidth cost data; According to the state vector, the action space, the immediate reward function, and the long-term reward function, an independent deep Q network is configured for each live broadcast room, wherein the deep Q network comprises four fully connected layers, the number of nodes in the four fully connected layers is 256, 128, 64, and 32 respectively, and the activation function of the deep Q network is the LeakyReLU function; Acquire time series differential error data, calculate sample priority based on the time series differential error data, and store the sample priority data in a playback buffer, wherein the playback buffer has a capacity of 10,000 and a sampling batch size of 64; An attention mechanism is constructed based on the sample data in the playback buffer, and the attention mechanism includes a query matrix, a key matrix, and a value matrix. The query matrix is ​​multiplied by the transposed matrix of the key matrix and divided by the square root of the dimension of the key matrix. The calculation result is input into the softmax function to obtain the attention weight, and the attention weight is multiplied by the value matrix to obtain the output result of multi-agent collaborative decision-making, so as to realize bandwidth allocation collaboration among multiple live broadcast rooms.

3. The method according to claim 1, characterized in that Based on the output results of the scenario classification model and the bandwidth allocation strategy, a multi-objective routing optimization model is constructed to dynamically allocate user access requests to the optimal edge node. The method also includes: A delay threshold of 50ms is set for the high-interaction game live broadcast, a freeze rate threshold of 0.1% is set for the education live broadcast, and a bit rate threshold of 8Mbps is set for the ordinary entertainment live broadcast, to build a service quality requirement matrix; Collect CPU utilization data, memory occupancy data, network bandwidth utilization data, and service quality index data of edge nodes, use a 5-minute sliding time window to calculate the stability scores of the CPU utilization data, the memory occupancy data, the network bandwidth utilization data, and the service quality index data, and obtain the comprehensive health score of the edge node through principal component analysis; Obtain user geographic location information data and network measurement data, calculate physical distance data, network hop count data, and historical delay data from the user to the edge node based on the user geographic location information data and the network measurement data, use a particle swarm optimization algorithm to adaptively adjust weight coefficients of the physical distance data, the network hop count data, and the historical delay data, and construct a user-edge node affinity matrix.

4. The method according to claim 3, characterized in that Based on the output results of the scenario classification model and the bandwidth allocation strategy, a multi-objective routing optimization model is constructed to dynamically allocate user access requests to the optimal edge nodes, including: A multi-objective routing optimization model is constructed based on the service quality requirement matrix, the comprehensive health score, and the user-edge node affinity matrix. The optimization objectives of the multi-objective routing optimization model include minimizing user access delay, maximizing edge node resource utilization, and minimizing cross-regional traffic costs. The constraints of the multi-objective routing optimization model include computing resource capacity constraints of edge nodes, network bandwidth capacity constraints, and service quality requirement constraints. An improved genetic algorithm is used to solve the multi-objective routing optimization model, the target edge node assigned by the user request is used as the gene position to construct the chromosome by integer coding, the chromosome is selected by the tournament selection method, the single point crossover method is used for crossover operation, and the uniform mutation method is used for mutation operation, wherein the improved genetic algorithm introduces a simulated annealing mechanism on the basis of the genetic algorithm to accept suboptimal solutions; A scheduling rule base is established based on the solution results of the multi-objective routing optimization model. The scheduling rule base contains scheduling rules for load fluctuations in different live broadcast scenarios. When a sudden load change is detected, an emergency scheduling mechanism is triggered to prioritize the service quality of high-priority live broadcast scenarios and dynamically allocate user access requests to the optimal edge node.

5. The method according to claim 4, characterized in that The method further includes a user requesting migration: The preheating resource scale of the target node is estimated based on historical load data, and the preheating time is dynamically adjusted through an exponential backoff algorithm. Computing resources are pre-allocated, cache is pre-heated, and network links are pre-established at the target node. The source node refers to the edge server node that is currently providing services to users, and the target node refers to the new edge server node selected to undertake service migration; Establishing a bidirectional data transmission channel between the source node and the target node, the bidirectional data transmission channel comprising an uplink channel for transmitting control information and a downlink channel for transmitting service data, adopting a three-way handshake mechanism to establish the bidirectional data transmission channel, monitoring the channel status of the bidirectional data transmission channel in real time, and switching to a standby channel when an abnormality is detected based on the channel status; Dynamic DNS resolution technology is used to direct newly generated user requests to the target node, a redirection ratio curve is constructed according to an S-shaped function, the number of requests directed to the target node is gradually increased according to the redirection ratio curve, the load state of the target node is monitored, and when the load state is abnormal, the user request is returned to the source node; The legacy request of the source node is managed for its life cycle, a maximum waiting time is set, the load status of the source node is monitored, and when the processing time of the legacy request exceeds the maximum waiting time, the legacy request is forcibly migrated to the target node through a compensation mechanism.

6. A bandwidth control and visitor traffic routing optimization system for multi-scenario live broadcast, used to implement the method described in any one of claims 1 to 5, characterized in that: include: The first unit is used to collect live broadcast data of multiple live broadcast scenes, wherein the live broadcast data includes network quality parameters, user interaction data and live broadcast content characteristics, wherein the network quality parameters include end-to-end delay, packet loss rate, and jitter, the user interaction data includes viewing time, interaction frequency, and barrage density, and the live broadcast content characteristics include video encoding parameters, audio bit rate, and picture complexity; Constructing a multi-dimensional feature vector based on the live broadcast data, using a deep learning algorithm to train a scene classification model, and dividing the live broadcast scenes into three types: high-interaction game live broadcast, ordinary entertainment live broadcast, and education live broadcast; The second unit is used to use the time series prediction algorithm to analyze the bandwidth demand change trend of various live broadcast scenarios and establish a bandwidth demand prediction model; according to the output results of the bandwidth demand prediction model and the total bandwidth resources of the live broadcast platform, a differentiated bandwidth allocation strategy is formulated to reserve sufficient bandwidth resources for high-interaction game live broadcasts and dynamically adjust bandwidth quotas for ordinary entertainment live broadcasts and educational live broadcasts; when fluctuations in bandwidth resources are detected, the bandwidth demand of high-interaction game live broadcasts is prioritized; The third unit is used to establish an edge node load status monitoring mechanism to collect computing resource occupancy, network throughput, and service quality indicators of each edge node in real time; Based on the output results of the scenario classification model and the bandwidth allocation strategy, a multi-objective routing optimization model is constructed to dynamically allocate user access requests to the optimal edge node; the multi-objective routing optimization model comprehensively considers the three goals of node load balance, user access delay, and bandwidth resource utilization, and uses a heuristic algorithm to solve the optimal routing solution to achieve real-time scheduling optimization of visitor traffic.

7. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 5.

8. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 5 is implemented.

Images