Network optimization method based on online conference

Abstract

The invention discloses a network optimization method based on an online conference, which relates to the technical field of real-time audio and video transmission, and comprises the following steps: in a transnational network, deploying monitoring probes at nodes participating in the online conference for monitoring network parameters such as network delay, bandwidth, packet loss rate and jitter in real time; the method comprises the following steps: processing network parameters and predicting abnormity by using a statistical analysis algorithm and a time sequence model to obtain network condition information, and then combining global network topology and using a reinforcement learning Q-Learning model and a graph neural network GNN topology prediction algorithm; according to the method, the network condition is monitored and predicted in real time, the optimal transmission path is output by using the reinforcement learning Q-Learning model and the GNN topology prediction algorithm, the bandwidth fluctuation is predicted through the LSTM, the packet loss rate is predicted based on the Bayesian network, the data transmission strategy is optimized, the network delay, the packet loss rate and the jitter are effectively reduced, and the network performance is improved. And the smoothness and the stability of the conference are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of real-time audio and video transmission, and particularly to a network optimization method based on online meetings. Background Art

[0002] In today's era of rapid digital and information development, significant progress has been made in real-time audio and video transmission technology. With the popularization of the Internet and the continuous improvement of bandwidth, real-time audio and video communication is no longer limited to traditional telephone conferences and video calls, but has been extended to many fields such as enterprise cross-border collaboration, distance education, online medical care, webcasting, and game interaction.

[0003] For example, in the distance education scenario, teachers teach students through real-time audio and video transmission, breaking geographical restrictions and realizing the sharing of high-quality educational resources; in the field of online medical care, doctors can communicate and diagnose with patients face-to-face through real-time audio and video, improving the efficiency and accessibility of medical services; webcasting has become a new form of entertainment and business, and the host conveys information and shows content to a large number of audiences through real-time audio and video; in game interaction, real-time voice communication between players also adds more fun to the game experience.

[0004] The existing technologies have the following deficiencies: Especially for online business meetings in enterprise cross-border collaboration, although the existing methods of real-time audio and video transmission technology can achieve basic communication, there are still many defects in meeting the requirements of low latency, high smoothness, and strong synchronization: For example, the long physical distance transmission in cross-border communication and the complexity of cross-border networks lead to high latency caused by frequent switching of transmission paths; the stability and reliability of cross-border network environments vary greatly, easily resulting in network congestion, network jitter, and packet loss problems, seriously affecting the quality of real-time audio and video transmission; and the clock drift existing across countries can cause audio-visual asynchronization, further affecting the stability and functional integrity of business meetings.

[0005] The above information disclosed in the background art section is only used to enhance the understanding of the background of the present disclosure, and thus it may include information that does not constitute the prior art known to those of ordinary skill in the art. Summary of the Invention

[0006] The object of the present invention is to provide a network optimization method based on online meetings. By using the reinforcement learning Q-Learning model and the graph neural network GNN topology prediction algorithm, the optimal transmission path is output. The bandwidth fluctuation is predicted by LSTM, and the packet loss rate is predicted based on the Bayesian network to optimize the data transmission strategy. By using the PTP algorithm to process the device local clock, the clock signal of GPS / Beidou satellite timekeeping is fused through Kalman filtering to correct the clock deviation, and the adaptive synchronization buffer ASB and the PID controller are used to dynamically adjust the audio and video playback timeline technology to solve the problems in the above background technology.

[0007] To achieve the above object, the present invention provides the following technical solutions: A network optimization method based on online meetings, including the following steps:

[0008] S1. In a cross-border network, monitoring probes are deployed at the nodes participating in the online meeting to monitor network parameters such as network latency, bandwidth, packet loss rate, and jitter in real time. The statistical analysis algorithm and the time series model are used to process and predict anomalies of the network parameters to obtain network condition information. Then, combined with the global network topology, the reinforcement learning Q-Learning model and the graph neural network GNN topology prediction algorithm are used to output the optimal transmission path;

[0009] S2. During the transmission of audio and video data along the optimal transmission path, LSTM is used to predict bandwidth fluctuations, the video resolution is dynamically adjusted through the contextual bandit model, and the improved weighted fair queue WFQ and random early detection RED are adopted to preferentially transmit packets with high priority. Based on the Bayesian network prediction of the packet loss rate, redundant packets are dynamically allocated according to the frame type, and a delay-constrained ARQ retransmission mechanism is set;

[0010] S3. The PTP algorithm is used to process the device local clock, the clock signal of GPS / Beidou satellite timekeeping is fused through Kalman filtering to correct the device clock deviation in real time, and the adaptive synchronization buffer ASB and the PID controller are used to dynamically adjust the audio and video playback timeline to synchronously control the audio and video streams and compensate for the decoding time difference;

[0011] S4. Based on compliance labels, the Kubernetes edge scheduler is used to dynamically deploy edge nodes. While ensuring that data is processed in compliance areas, after the audio and video data of the online meeting is processed by the H.264 / H.265 video encoding and Opus audio encoding algorithms, the ROI-driven compression and caching decision model is combined with the international bandwidth cost to dynamically select the compression ratio and pre-cache meeting materials;

[0012] S5. Collect user feedback data and index data that affect the average latency, packet loss rate, and audio-visual synchronization error of the meeting performance. Use the Deep Deterministic Policy Gradient (DDPG) algorithm to dynamically adjust the routing and network parameters according to the real-time performance reward value of the meeting.

[0013] Optionally, the steps for obtaining the network condition information are as follows:

[0014] Deploy monitoring probes at n nodes of the cross-border network to collect network parameters such as network latency, bandwidth, packet loss rate, and jitter in real time, and calibrate them as NP n ={D, B, L, J} n , and the units corresponding to the network latency D, bandwidth B, packet loss rate L, and jitter J are ms (milliseconds), Mbps (megabits per second), %, and ms (milliseconds);

[0015] The statistical analysis algorithm calculates the average value and standard deviation of the network parameters NP and uses the Z-score algorithm for anomaly detection. Among them, the formula for the average value is and i ∈ n, where μ NP represents the average value of the network parameters NP, n represents the total number of nodes in the cross-border network, i represents the i-th node in the total number of nodes, and NP i represents the network parameter value collected corresponding to the i-th node;

[0016] The formula for the standard deviation is where σ NP represents the standard deviation of the network parameters NP;

[0017] The expression of the Z-score algorithm is When , represents the network parameter NP collected corresponding to the i-th node i is determined to be abnormal;

[0018] Use the Autoregressive Integrated Moving Average (ARIMA) model for time series prediction to predict the bandwidth trend of the next sliding window. Among them, the formula for the ARIMA model to predict the bandwidth trend is B t+1 = φ1B t + θ1∈ t + ∈ t+1 , where B t+1 represents the bandwidth value predicted by the ARIMA model at time t + 1, B t represents the bandwidth value of the time series of the bandwidth parameter at the current time t, φ1 represents the autoregressive coefficient, θ1 represents the moving average coefficient, t represents the error term of the white noise at the current time t, t+1 represents the error term of the white noise at time t + 1;

[0019] Generate network condition information by extracting the characteristic information of bandwidth volatility, jitter accumulation value, and packet loss rate trend, and use the network condition information as the input of the subsequent model. Among them, the calculation formula of bandwidth volatility is and \(t\in T\). In the formula, \(T\) represents the size of the sliding window, \(\Delta B\) represents the bandwidth volatility, and \(B\) {1,2，…，T} represents the bandwidth sequence within the time window \(T\), \(max(B\) {1,2，…，T} ), \(min(B\) {1,2，…，T} ), and \(\mu(B\) {1,2，...，T} respectively represent the maximum value, minimum value, and average value of the bandwidth sequence \(B\) within the time window \(T\); {1,2，…，T}

[0020] The calculation formula of the jitter accumulation value is In the formula, \(J'\) represents the jitter accumulation value, and \(J\) t represents the jitter value at time \(t\), and \(J\) t-1 represents the initial jitter value, and \(J\) t-1 = \(0\);

[0021] The calculation formula of the packet loss rate trend is and In the formula, \(\kappa(L)\) represents the packet loss rate trend, \(t\) represents the mean value of the time index within the time window \(T\), and \(\mu\) L represents the mean value of the packet loss rate within the time window \(T\);

[0022] The expression of the network condition information is \(NCI = (\Delta B, J', \kappa(L))\), where \(NCI\) represents the network condition information.

[0023] Optionally, the output steps of the optimal transmission path are as follows:

[0024] Define the state space and action space in reinforcement learning according to network parameters. Among them, the expression of the state space is \(S_s=\{D, B, L, J, C_m\}\), where \(S_s\) represents the state space of inserting the compliance mark \(C_m\) of network parameters, and the dimension is \(5\times1\). \(C_m\) represents the compliance mark of whether the path violates data sovereignty, and the boolean value of \(C_m\) is \(0\) or \(1\);

[0025] The expression of the action space is \(A_s = \{Switch the current path, Keep the current path\}\), where \(A_s\) represents the action space;

[0026] ​Design a multi-objective reward function and use the Q-Learning algorithm for iterative learning to update the Q-value table. Among them, the expression of the multi-objective reward function is R = -(αD + βL + γCm), and α + β + γ = 1, Cm = 0. In the formula, R represents the reward function, and α, β, and γ represent the weight coefficients corresponding to the network delay D, packet loss rate L, and non-compliance marker Cm of violating data sovereignty;

[0027] The calculation formula for updating the Q value is In the formula, Q′(Ss,As) represents the updated Q value, Q(Ss,As) represents the Q value of taking action As in state Ss, η represents the learning rate, and η ∈ (0, 1], λ represents the discount factor, represents the maximum Q value among all actions As in the next state Ss;

[0028] Abstract the global network topology into a graph structure. Nodes represent devices such as routers and servers in the network, and edges represent the links connecting the devices. Then, the attributes of each edge are the network delay D and bandwidth B. Among them, the expression of the graph structure is Gs = (v i , e j ). In the formula, v i represents the nodes of network devices, and e j represents the attributes of the link;

[0029] In the graph structure, according to the message passing mechanism of audio-visual data in the GNN topology, use the graph neural network GNN to predict the link congestion probability, and combine the action selection of reinforcement learning Q-Learning to output the optimal transmission path. Among them, the calculation formula for the link congestion probability is Pc = σ(ω P ·h(e j ) + b P ), and Pc ∈ [0, 1]. In the formula, Pc represents the link congestion probability, σ represents the Sigmoid activation function, ω P represents the weight coefficient of the link congestion prediction layer, h(e j ) represents the embedding of the link attribute e j in the GNN message passing mechanism, and b P represents the bias term of the link congestion prediction layer;

[0030] The expression for the action selection of Q-Learning is In the formula, As” represents the action selected by Q-Learning, and argmax() represents the argmax function for taking the maximum value;

[0031] The output formula for the optimal transmission path is In the formula, p* represents the optimal transmission path, Denoted as the set of all feasible paths in the graph structure, argmin() represents the argmin function for minimization, and D(e j ) represents the delay of the link, φ represents the indicator function for the link congestion probability Pc, and Pc > 0.7 represents high link congestion.

[0032] Optionally, the steps for the context bandit model to adjust the video resolution are as follows:

[0033] Using the context bandit model, predict the bandwidth fluctuation from the network condition information NCI using LSTM, and extract the real-time network state features and calibrate them as nsc according to the current bandwidth, current network jitter, and the priority of the defined frame types t =[ΔB t , B t , J t , FP t T , where nsc t represents the network state features at the current time t, ΔB t represents the bandwidth change rate predicted by LSTM for a small period in the future, B t represents, J t represents, FP t represents the frame priority at the current time t, and the frame types include I frames, P frames, and B frames, and it is defined that the priority of I frames is higher than that of P frames, which is higher than that of B frames, that is, FP t =(FP(I)>FP(P)>FP(B))t, where T represents the transpose;

[0034] Select a resolution to define the bitrate level Design the reward function for the resolution, balancing the picture quality and stuttering. Among them, the expression of the reward function for the resolution is and ω1 + ω2 + ω3 = 1. Among them, represents the reward function for the resolution, ω1 represents the reward weight coefficient corresponding to the picture quality, represents the picture quality, ω2 represents the penalty weight coefficient corresponding to stuttering, represents stuttering, and has a value of 0 or 1, ω3 represents the penalty weight coefficient corresponding to bandwidth matching, represents bandwidth matching;

[0035] ​Packets are preferentially queued and sent according to the frame priority FP using improved weighted fair queueing (WFQ) and random early detection (RED), and the packet loss threshold is dynamically adjusted. In combination with the ε-greedy strategy, the coding rate level is selected to determine the optimal transmission video resolution. The calculation formula for the optimal transmission video resolution is In the formula,[[]] represents the optimal transmission video resolution,[[]] represents the coding rate level[[]] Value function estimation.[[]]

[0036] Optionally, the steps of the PTP algorithm processing device are as follows:[[]]

[0037] In the network, determine the master clock and slave clock of the device. First, use the master clock to send a synchronization Sync message and record the master clock transmission timestamp as

[0038] After the slave clock receives the synchronization Sync message, record the reception timestamp as

[0039] The slave clock sends a delay request Delay_Req message to the master clock and records the slave clock transmission timestamp as

[0040] After the master clock receives the delay request Delay_Req message, record the master clock reception timestamp as Send to the slave clock;[[]]

[0041] The slave clock calculates the clock deviation and network delay between itself and the master clock based on the received timestamps. Among them, the calculation formula for the clock deviation is In the formula, Offset represents the clock deviation of the slave clock relative to the master clock;[[]]

[0042] The calculation formula for the network delay is In the formula, Delay represents the network delay between the master and slave clocks;[[]]

[0043] The slave clock adjusts its local clock according to the calculated clock deviation to synchronize the slave clock with the master clock.[[]]

[0044] Optionally, the correction steps for the device clock deviation are as follows:[[]]

[0045] While collecting the clock signal of GPS / Beidou satellite time service, obtain the clock deviation and frequency deviation between the local clock and the GPS / Beidou satellite reference clock, define the clock state vector and calculate the clock state equation simultaneously. The expression of the clock state vector is In the formula, x kDenoted as the clock state vector, Δt k Denoted as the clock offset, Δf k Denoted as the frequency offset;

[0046] The calculation formula of the clock state equation is In the formula, x k+1 Denoted as the clock state at the next moment, Denoted as the state transition matrix, and T′ is denoted as the filtering period, w k Denoted as the process noise subject to Gaussian distribution;

[0047] Based on the PTP algorithm, an observation equation is constructed by fusing the local clock time deviation and the time deviation measured by the GPS / Beidou satellite signal, and the Kalman filtering algorithm is used for iterative prediction and update to obtain the clock deviation estimated value of the posterior state. Among them, the expression of the observation equation is z k =[1, 0]x k +v k , and Δt k =Δt k PTP +Δt k GPS In the formula, z k Denoted as the observation equation, [1, 0] is denoted as the observation matrix, v k Denoted as the observation noise subject to Gaussian distribution, Δt k PTP Denoted as the local clock time deviation obtained by the PTP algorithm, Δt k GPS Denoted as the time deviation measured by the GPS / Beidou satellite signal;

[0048] Perform real-time correction and adjustment of the time and frequency of the local clock according to the clock deviation estimated value of the posterior state.

[0049] Optionally, the processing steps of the ROI-driven compression and caching decision model are as follows:

[0050] Construct a bandwidth cost model according to the dynamic change of the international bandwidth cost with the link and time period. Among them, the expression of the bandwidth cost model is In the formula, C BW Denoted as the total bandwidth cost, Volume t Denoted as the transmission data volume in time period t, Price t Denoted as the bandwidth unit price in time period t, T0 is denoted as the total time period length;

[0051] After processing the audio and video data of the online meeting using the H.264 / H.265 video encoding and Opus audio encoding algorithms, the resulting compressed data will reduce the image quality. The peak signal-to-noise ratio (PSNR) is used to quantify the loss value of the image quality. Among them, the calculation formula for the loss value of the image quality is In the formula, ΔPQ represents the loss of image quality after processing the audio and video data, MSE represents the mean square error of the image after compressing the audio and video data, 255 represents the maximum pixel value, and PSNR0 represents the PSNR value of the original audio and video;

[0052] Combining the bandwidth cost and the loss value of the image quality to calculate the return on investment (ROI), and obtaining the compression ratio decision rule according to the constraints of the ROI-driven compression and caching decision model. Among them, the calculation formula for the return on investment (ROI) is And In the formula, ROI represents the return on investment, ΔC BW Represents the bandwidth cost savings value, ζ represents the image quality loss weight coefficient, ζ·ΔPQ represents the image quality loss cost savings value, TC represents the total cost of bandwidth and image quality, and CR represents the compression ratio;

[0053] The expression of the compression ratio decision rule is In the formula, CR′ represents the maximum compression ratio;

[0054] And formulating a cache revenue target through the ROI-driven compression and caching decision model to obtain the cache decision rule, which is used to pre-cache high-frequency meeting materials to reduce the cost of repeated transmission. Among them, the calculation formula for the cache decision rule is CC m = p m ·s m ·C BW , In the formula, CC m Represents the expected revenue of caching the m-th meeting material, p m Represents the access probability of the m-th meeting material, s m Represents the size of the storage space occupied by the m-th meeting material, and m represents the number of pre-cached meeting materials.

[0055] Optionally, the dynamic adjustment logic of the Deep Deterministic Policy Gradient (DDPG) algorithm is as follows:

[0056] Collect user feedback data and metric data affecting meeting performance, including average latency, packet loss rate, and audio-visual synchronization error;

[0057] Integrate the collected metric data to define the state of the current meeting;

[0058] Design a reward function to calculate the reward value according to the real-time performance state of the meeting;

[0059] Using the policy network and value network in the DDPG algorithm, the policy network outputs an action according to the current state, and the policy network evaluates the value of taking the action in the state. Among them, the actions include routing switching, FEC redundancy, and bit rate adjustment;

[0060] Store each state, action, reward, and next state into the experience replay buffer;

[0061] Randomly sample a batch of data from the experience replay buffer, and define the loss function of the policy network to update the parameters of the policy network and value network;

[0062] According to the updated policy network, select the optimal action in the new state and apply it to the conferencing system to dynamically adjust the routing and network parameters;

[0063] Iterate the steps of the DDPG algorithm, continuously collect data, calculate rewards, update network parameters, and adjust routing and network parameters.

[0064] Optionally, a computer device includes: a memory and a processor; the memory stores a computer program, and when the processor executes the computer program, the steps of the above-mentioned network optimization method based on online conferencing are implemented.

[0065] Optionally, a computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the above-mentioned network optimization method based on online conferencing are implemented.

[0066] In the above technical solution, the technical effects and advantages provided by the present invention:

[0067] The present invention monitors and predicts the network condition in real time, uses the reinforcement learning Q-Learning model and the graph neural network GNN topology prediction algorithm to output the optimal transmission path, predicts the bandwidth fluctuation through LSTM and predicts the packet loss rate based on the Bayesian network, optimizes the data transmission strategy, effectively reduces the network latency, packet loss rate and jitter, and improves the fluency and stability of the conference; by using the PTP algorithm to process the device local clock, combining the Kalman filter to fuse the clock signals of GPS / Beidou satellite time synchronization to correct the clock deviation, and using the adaptive synchronization buffer ASB and the PID controller to dynamically adjust the audio-visual playback timeline technology, it not only compensates the decoding time difference in real time to ensure high synchronization of audio and video, but also enhances the user's audio-visual experience; based on the compliance label, the edge nodes are dynamically deployed to ensure that the data is processed in the area that complies with local regulations, avoiding legal risks, and combining the ROI-driven compression and caching decision model and the international bandwidth cost, dynamically select the compression ratio and pre-cache the conference materials, while ensuring the conference quality, reducing the bandwidth cost; by collecting user feedback and conference performance metric data, using the DDPG algorithm to dynamically adjust the routing and network parameters in real time, realizing the long-term optimization of conference performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required in the embodiments. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those of ordinary skill in the art, other drawings can also be obtained based on these drawings.

[0069] Figure 1 It is a flowchart of the network optimization method based on online conferencing of the present invention. DETAILED DESCRIPTION

[0070] Now, the example embodiments will be described more fully with reference to the accompanying drawings. However, the example embodiments can be implemented in various forms and should not be construed as limited to the examples set forth herein; rather, these example embodiments are provided so that this disclosure will be more comprehensive and complete, and will fully convey the concept of the example embodiments to those skilled in the art.

[0071] The present invention provides a network optimization method based on online conferencing as shown in Figure 1 the following, including the following steps:

[0072] S1. In a transnational network, monitoring probes are deployed at nodes participating in online meetings to monitor network parameters such as network latency, bandwidth, packet loss rate, and jitter in real time. Statistical analysis algorithms and time series models are used to process and predict anomalies in the network parameters to obtain network condition information. Then, combined with the global network topology, the reinforcement learning Q-Learning model and the graph neural network GNN topology prediction algorithm are used to output the optimal transmission path with low latency and high stability for transmitting the audio and video data of the meeting.

[0073] Specifically, the steps for obtaining network condition information are as follows:

[0074] Monitoring probes are deployed at n nodes in the transnational network to collect network parameters such as network latency, bandwidth, packet loss rate, and jitter in real time, and are respectively calibrated as NP n ={D, B, L, J} n , and the units corresponding to network latency D, bandwidth B, packet loss rate L, and jitter J are ms (milliseconds), Mbps (megabits per second), %, and ms (milliseconds);

[0075] The statistical analysis algorithm calculates the average value and standard deviation of the network parameters NP and uses the Z-score algorithm for anomaly detection. Among them, the calculation formula for the average value is and i ∈ n. In the formula, μ NP represents the average value of the network parameters NP, n represents the total number of nodes in the transnational network, i represents the i-th node in the total number of nodes, and NP i represents the network parameter value collected corresponding to the i-th node;

[0076] The calculation formula for the standard deviation is In the formula, σ NP represents the standard deviation of the network parameters NP;

[0077] The expression of the Z-score algorithm is When , represents the network parameter NP collected corresponding to the i-th node i is determined to be abnormal;

[0078] The time series prediction ARIMA model is used to predict the bandwidth trend of the sliding window at the next moment. Among them, the calculation formula for the ARIMA model to predict the bandwidth trend is B t+1 =φ1B t +θ1∈ t +∈ t+1 , in the formula, B t+1 represents the bandwidth value predicted by the ARIMA model at the t + 1 moment, and B tIt is represented as the bandwidth value of the time series of bandwidth parameters at the current time t, φ1 is represented as the autoregressive coefficient, θ1 is represented as the sliding average coefficient, t It is expressed as the error term of white noise at the current time t, t+1 It is represented as the error term of white noise at time t+1;

[0079] The network status information is generated by extracting the characteristic information of bandwidth fluctuation rate, jitter accumulation value and packet loss rate trend, and the network status information is used as the input of the subsequent model. The calculation formula of bandwidth fluctuation rate is: And t∈T, where T represents the size of the sliding window, ΔB represents the bandwidth fluctuation rate, and B {1,2，…，T} It is expressed as the bandwidth sequence in the time window T, max(B {1,2，…，T} )、min(B {1,2，…，T} ), μ(B {1,2，...，T} ) are respectively represented as the bandwidth sequence B in the time window T {1,2，…，T} the maximum, minimum and average values of ;

[0080] The calculation formula for the jitter accumulation value is: Where J' is the jitter accumulation value, J t Expressed as the jitter value at time t, J t-1 is the initial jitter value, and J t-1 =0;

[0081] The calculation formula for packet loss rate trend is: and Where κ(L) represents the packet loss rate trend, t represents the mean of the time index in the time window T, and μ L Expressed as the mean packet loss rate within the time window T;

[0082] The expression of network status information is NCI=(ΔB, J', κ(L)), where NCI represents network status information.

[0083] Specifically, the steps for outputting the optimal transmission path are as follows:

[0084] Define the state space and action space in reinforcement learning based on network parameters. The state space is expressed as Ss = {D, B, L, J, Cm}, where Ss represents the state space of network parameters with the compliance marker Cm inserted, and its dimension is 5 × 1. Cm represents the compliance marker indicating whether the path violates data sovereignty, and Cm has a Boolean value of 0 or 1.

[0085] The expression of action space is As = {switch current path, maintain current path}, where As represents the action space;

[0086] Design a multi-objective reward function and use the Q-Learning algorithm for iterative learning to update the Q-value table. Among them, the expression of the multi-objective reward function is R = -(αD + βL + γCm), and α + β + γ = 1, Cm = 0. In the formula, R represents the reward function, and α, β, and γ represent the weight coefficients corresponding to the network delay D, packet loss rate L, and non-compliance marker Cm of violating data sovereignty;

[0087] The calculation formula for updating the Q value is In the formula, Q′(Ss,As) represents the updated Q value, Q(Ss,As) represents the Q value of taking action As in state Ss, η represents the learning rate, and η ∈ (0, 1], λ represents the discount factor, represents the maximum Q value among all actions As in the next state Ss;

[0088] Abstract the global network topology into a graph structure. Nodes represent devices such as routers and servers in the network, and edges represent the links connecting the devices. Then, the attributes of each edge are network delay D and bandwidth B. Among them, the expression of the graph structure is Gs = (v i , e j ). In the formula, v i represents the nodes of network devices, and e j represents the attributes of the links;

[0089] In the graph structure, according to the message passing mechanism of audio-visual data in the GNN topology, use the graph neural network GNN to predict the link congestion probability, and combine the action selection of reinforcement learning Q-Learning to output the optimal transmission path. Among them, the calculation formula for the link congestion probability is Pc = σ(ω P ·h(e j ) + b P ), and Pc ∈ [0, 1]. In the formula, Pc represents the link congestion probability, σ represents the Sigmoid activation function, ω P represents the weight coefficient of the link congestion prediction layer, h(e j ) represents the embedding of the link attribute e j in the GNN message passing mechanism, and b P represents the bias term of the link congestion prediction layer;

[0090] The expression for the action selection of Q-Learning is In the formula, As” represents the action selected by Q-Learning, and argmax() represents the argmax function for taking the maximum value;

[0091] The output formula for the optimal transmission path is In the formula, p* represents the optimal transmission path, Denoted as the set of all feasible paths in the graph structure, argmin() denotes taking the minimum of the argmin function, D(e j ) denotes the delay of the link, φ denotes the indicator function for the link congestion probability Pc, and Pc > 0.7 denotes high link congestion.

[0092] Furthermore, the transnational network is complex and changeable. By real-time monitoring and data analysis, and then grasping the dynamic changes of the transnational network, it is an effective path planning to provide a reliable basis for subsequent path selection and optimization. Using the reinforcement learning Q-Learning model and the graph neural network GNN topology prediction algorithm, combined with the global network topology and real-time network conditions, it can quickly output the optimal transmission path with low latency and high stability. Effectively reducing the latency and jitter of audio and video transmission is a crucial processing method, and then solving problems such as unclear network conditions, poor path selection, and insufficient response to anomalies, significantly improving the network quality of transnational online meetings, ensuring the fluency, stability, and audio-video synchronization of the meetings, and providing a high-quality meeting experience for participants.

[0093] S2. During the transmission of audio and video data along the optimal transmission path, use LSTM to predict bandwidth fluctuations, dynamically adjust the video resolution through the contextual bandit model, and adopt the improved weighted fair queue WFQ and random early detection RED to preferentially transmit packets with high priority, and predict the packet loss rate based on the Bayesian network, dynamically allocate redundant packets according to the frame type, and set the delay-constrained ARQ retransmission mechanism to refresh the packet tags for key frames to ensure the stable transmission of audio and video data and avoid packet loss in network transmission;

[0094] Specifically, the steps for the contextual bandit model to adjust the video resolution are as follows:

[0095] Use the contextual bandit model to predict the bandwidth fluctuation from the network condition information NCI using LSTM, and extract the real-time network state features and calibrate them as nsc according to the current bandwidth, current network jitter, and the priority of the defined frame type t =[ΔB t ,B t ,J t ,FP t T where nsc t denotes the network state feature at the current time t, ΔB t denotes the bandwidth change rate predicted by LSTM for a small future time period, B t denotes, J t denotes, FP t ​Denoted as the frame priority at the current time t, and the frame types include I-frame, P-frame, and B-frame, and it is defined that the priority of the I-frame is higher than that of the P-frame, and the priority of the P-frame is higher than that of the B-frame, that is, FP t =(FP(I) > FP(P) > FP(B)) t , T is denoted as transpose;

[0096] Select the resolution to define the bitrate level {α1: 4Mbps(1080p), α2: 3Mbps(720p), α3: 2Mbps(480p)}, design the reward function of the resolution, balance the picture quality and stuttering. Among them, the expression of the reward function of the resolution is And ω1 + ω2 + ω3 = 1. In the formula, Denoted as the reward function of the resolution, ω1 is denoted as the reward weight coefficient corresponding to the picture quality, Denoted as the picture quality, ω2 is denoted as the penalty weight coefficient corresponding to the stuttering, Denoted as the stuttering, and The value of is 0 or 1, ω3 is denoted as the penalty weight coefficient corresponding to the bandwidth matching, Denoted as the bandwidth matching;

[0097] According to the frame priority FP, use the improved weighted fair queue WFQ and random early detection RED to prioritize the sending of data packets, and dynamically adjust the packet loss threshold. Cooperate with the ε-greedy strategy to select the bitrate level and determine the optimal transmission video resolution. Among them, the calculation formula of the optimal transmission video resolution is In the formula, Denoted as the optimal transmission video resolution, Denoted as the bitrate level Value function estimation.

[0098] Specifically, the steps of predicting the packet loss rate by the Bayesian network are as follows:

[0099] Fuse the prior knowledge of geographical location, operator, and transmission protocol, define the root nodes of the Bayesian network, describe the network congestion level according to the link congestion probability Pc as the intermediate nodes of the Bayesian network, and then combine the packet loss rate L defined as the leaf nodes of the Bayesian network;

[0100] Construct a conditional probability model according to the historical network congestion level and packet loss rate distribution data, obtain the predicted packet loss rate and dynamically adjust the FEC redundancy, and use the FEC redundancy to weight the frame types to calculate the data packets with high transmission priority.

[0101] Furthermore, by predicting bandwidth fluctuations through LSTM and dynamically adjusting the video resolution using the contextual bandit model, the audio-visual data transmission can be matched with the real-time bandwidth conditions, avoiding lags caused by insufficient bandwidth. Meanwhile, when the bandwidth is sufficient, high video quality can be ensured. The improved WFQ and RED algorithms are used to ensure that high-priority data packets are transmitted first, guaranteeing the timely delivery of key audio-visual data, improving the fluency and stability of the meeting, and reducing the loss and delay of important information. In addition, predicting the packet loss rate based on the Bayesian network and dynamically allocating redundant packets according to the frame type, as well as setting up a delay-constrained ARQ retransmission mechanism, can more precisely address the packet loss problem. While ensuring data integrity, it improves the transmission efficiency and reduces unnecessary retransmissions and delays.

[0102] S3. Use the PTP algorithm to process the device's local clock, fuse the clock signals of GPS / Beidou satellite timekeeping through Kalman filtering, correct the device clock deviation in real time, and use the adaptive synchronization buffer ASB and PID controller to dynamically adjust the audio-visual playback timeline to synchronously control the audio-visual stream, compensate for the decoding time difference, solve the cross-country clock drift, and ensure audio-visual synchronization.

[0103] Specifically, the steps of using the PTP algorithm to process the device are as follows:

[0104] In the network, determine the master clock and slave clock of the device. First, use the master clock to send a synchronization Sync message and record the master clock sending timestamp as

[0105] After the slave clock receives the synchronization Sync message, record the reception timestamp as

[0106] The slave clock sends a delay request Delay_Req message to the master clock and records the slave clock sending timestamp as

[0107] After the master clock receives the delay request Delay_Req message, record the master clock reception timestamp as Send to the slave clock;

[0108] The slave clock calculates the clock deviation and network delay between itself and the master clock according to the received timestamps. Among them, the calculation formula for the clock deviation is In the formula, Offset represents the clock deviation of the slave clock relative to the master clock;

[0109] The calculation formula for the network delay is In the formula, Delay represents the network delay between the master and slave clocks;

[0110] The slave clock adjusts the local clock according to the calculated clock deviation so that the slave clock is synchronized with the master clock.

[0111] Specifically, the steps for correcting the device clock deviation are as follows:

[0112] While collecting the clock signal of GPS / Beidou satellite time service, obtain the clock deviation and frequency deviation between the local clock and the GPS / Beidou satellite reference clock, define the clock state vector and calculate the clock state equation. Among them, the expression of the clock state vector is In the formula, x k represents the clock state vector, Δt k represents the clock deviation, and Δf k represents the frequency deviation;

[0113] The calculation formula of the clock state equation is In the formula, x k+1 represents the clock state at the next moment, represents the state transition matrix, and T′ represents the filtering period, and w k represents the process noise subject to Gaussian distribution;

[0114] Based on the PTP algorithm, construct an observation equation by fusing the local clock time deviation and the time deviation measured by the GPS / Beidou satellite signal, and use the Kalman filtering algorithm for iterative prediction and update to obtain the estimated value of the clock deviation of the posterior state. Among them, the expression of the observation equation is z k =[1, 0]x k +v k , and Δt k =Δt k PTP +Δt k GPS , in the formula, z k represents the observation equation, [1, 0] represents the observation matrix, and v k represents the observation noise subject to Gaussian distribution, and Δt k PTP represents the local clock time deviation obtained by the PTP algorithm, and Δt k GPS represents the time deviation measured by the GPS / Beidou satellite signal;

[0115] Perform real-time correction and adjustment of the time and frequency of the local clock according to the estimated value of the clock deviation of the posterior state.

[0116] Specifically, the steps for synchronizing and controlling the audio and video stream are as follows:

[0117] During the acquisition and encoding of audio and video data, add timestamps to each frame of audio and video data and calculate the decoding time difference of the audio and video.

[0118] Audio and video data enters the adaptive synchronization buffer ASB. The buffer dynamically adjusts the buffer depth according to the timestamps of the audio and video and the current playback progress to compensate for the decoding time difference.

[0119] Use a PID controller to calculate the adjustment amount according to the time difference between the audio and video, and dynamically adjust the playback timeline of the audio and video to keep the audio and video streams synchronized.

[0120] And calculate the root mean square of the synchronization error to monitor and feedback the synchronization of the audio and video streams in real time.

[0121] Furthermore, by fusing the satellite timing signal through the PTP algorithm and Kalman filter, the clock deviation of the device can be corrected in real time and accurately, so that the clocks of different devices are highly synchronized, providing a solid time basis for the synchronous transmission of audio and video; through the collaborative work of the adaptive synchronization buffer ASB and the PID controller, the playback timeline can be dynamically adjusted according to the actual decoding situation and time difference of the audio and video, effectively compensating for the decoding time difference, ensuring that the audio and video streams are always synchronized during playback, not only solving the problems of cross-border clock drift and audio-video asynchronization, but also significantly improving the audio and video quality in scenarios such as cross-border online meetings and cross-border video live broadcasts, bringing users a smooth, clear and synchronous audio and video experience.

[0122] S4. Based on the compliance label, use the Kubernetes edge scheduler to dynamically deploy edge nodes. While ensuring that data is processed in the compliance area, after processing the audio and video data of the online meeting using the H.264 / H.265 video encoding and Opus audio encoding algorithms, use the ROI-driven compression and caching decision model combined with the international bandwidth cost to dynamically select the compression ratio and pre-cache the meeting materials.

[0123] Specifically, the steps for the dynamic deployment of edge nodes are as follows:

[0124] Based on the compliance mark that the path does not violate data sovereignty Cm = 1, label each edge node.

[0125] The Kubernetes edge scheduler collects information about each edge node, including the resource metrics of the node's CPU / memory and the network status of network latency / bandwidth.

[0126] According to the collected information of the edge nodes and the requirements of the compliance mark, formulate a scheduling policy, and the goal of the policy is to deploy data and application programs to edge nodes that meet the compliance requirements and have appropriate resources.

[0127] According to the scheduling policy, the Kubernetes edge scheduler dynamically deploys data and application programs to the selected edge nodes.

[0128] Specifically, the processing steps of the ROI-driven compression and caching decision model are as follows:

[0129] Construct a bandwidth cost model according to the dynamic change of international bandwidth cost with the link and time period. Among them, the expression of the bandwidth cost model is In the formula, C BW represents the total bandwidth cost, Volume t represents the amount of transmitted data in time period t, Price t represents the unit price of bandwidth in time period t, and T0 represents the total time period length;

[0130] After processing the audio and video data of the online meeting using the H.264 / H.265 video coding and Opus audio coding algorithms, the resulting compressed data will reduce the image quality. The peak signal-to-noise ratio PSNR is used to quantify the loss value of the image quality. Among them, the calculation formula for the loss value of the image quality is In the formula, ΔPQ represents the loss of image quality after processing the audio and video data, MSE represents the mean square error of the image after compressing the audio and video data, 255 represents the maximum pixel value, and PSNR0 represents the PSNR value of the original audio and video;

[0131] Calculate the return on investment ROI by combining the bandwidth cost and the loss value of the image quality, and obtain the compression ratio decision rule according to the constraint conditions defined by the ROI-driven compression and caching decision model. Among them, the calculation formula for the return on investment ROI is And In the formula, ROI represents the return on investment, ΔC BW represents the saved value of the bandwidth cost, ζ represents the weight coefficient of the image quality loss, ζ·ΔPQ represents the saved value of the image quality loss cost, TC represents the total cost of the bandwidth and the image quality, and CR represents the compression ratio;

[0132] The expression of the compression ratio decision rule is In the formula, CR′ represents the maximum compression ratio;

[0133] And formulate a cache revenue target through the ROI-driven compression and caching decision model to obtain the cache decision rule, which is used to pre-cache high-frequency meeting materials to reduce the cost of repeated transmission. Among them, the calculation formula for the cache decision rule is CC m = p m ·s m ·C BW , where CC m represents the expected revenue of caching the mth meeting material, p m represents the access probability of the mth meeting material, s m represents the size of the storage space occupied by the mth meeting material, and m represents the number of pre-cached meeting materials.

[0134] Furthermore, by dynamically deploying edge nodes according to compliance tags through the Kubernetes edge scheduler, it ensures that audio and video data is strictly processed within the compliant area, effectively avoiding legal risks and business risks caused by non-compliant data processing; adopting H.264 / H.265 video coding and Opus audio coding algorithms significantly improves the data compression rate while ensuring the quality of audio and video, reducing the resources required for data transmission and storage; and by combining the ROI-driven compression and caching decision model with international bandwidth costs for dynamic decision-making, it can flexibly select the optimal compression rate and pre-caching strategy under different bandwidth conditions, effectively controlling the bandwidth cost while ensuring the smoothness and stability of online meetings and enhancing the overall meeting experience.

[0135] S5. Collect user feedback data and metric data affecting the average latency, packet loss rate, and audio-visual synchronization error of the meeting performance. Use the Deep Deterministic Policy Gradient (DDPG) algorithm to dynamically adjust routing and network parameters according to the real-time performance reward value of the meeting to achieve long-term performance improvement.

[0136] Specifically, the dynamic adjustment logic of the Deep Deterministic Policy Gradient (DDPG) algorithm is as follows:

[0137] Collect user feedback data and metric data affecting the meeting performance, including average latency, packet loss rate, and audio-visual synchronization error;

[0138] Integrate the collected metric data to define the current state of the meeting;

[0139] Design a reward function to calculate the reward value according to the real-time performance state of the meeting;

[0140] Use the policy network and value network in the DDPG algorithm. The policy network outputs an action according to the current state, and the policy network evaluates the value of taking the action in the state. Among them, the actions include routing switching, FEC redundancy, and bitrate adjustment;

[0141] Store each state, action, reward, and next state in the experience replay buffer;

[0142] Randomly sample a batch of data from the experience replay buffer and define the loss function of the policy network to update the parameters of the policy network and value network;

[0143] According to the updated policy network, select the optimal action in the new state and apply it to the meeting system to dynamically adjust routing and network parameters;

[0144] Steps of the iterative DDPG algorithm, continuously collect data, calculate rewards, update network parameters, and adjust routing and network parameters to achieve long-term improvement in meeting performance.

[0145] Furthermore, the use of the DDPG algorithm enables the system to quickly adjust routing and network parameters according to real-time collected user feedback and performance metric data to adapt to the changing network environment, ensuring that the meeting can maintain good performance in various situations; through continuous data collection and dynamic adjustment, the meeting performance can be continuously optimized, reducing the average latency, packet loss rate, and audio-visual synchronization error. In the long run, the quality and stability of cross-border online meetings are significantly improved; stable and high-quality meeting performance can effectively reduce problems such as stuttering and out-of-sync audio and video encountered by users during the meeting, thereby enhancing user satisfaction and participation in the meeting.

[0146] The software automation vulnerability detection system based on fuzz testing provided by the embodiments of the present invention is implemented through the above-mentioned network optimization method based on online meetings. The specific methods and processes of the software automation vulnerability detection system based on fuzz testing can be found in the embodiments of the above-mentioned network optimization method based on online meetings, and will not be elaborated here.

[0147] The above formulas are all dimensionless and take their numerical values for calculation. The formula is obtained by collecting a large amount of data for software simulation to obtain a formula that is closest to the real situation. The preset parameters in the formula are set by those skilled in the art according to the actual situation.

[0148] The above embodiments can be implemented in whole or in part by software, hardware, firmware, or any other combination. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the processes or functions described in the embodiments of the present application are generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. 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, computer, server, or data center to another website, computer, server, or data center by wired or wireless (such as infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or data center that contains one or more collections of available media. The available medium can be a magnetic medium (such as a floppy disk, hard disk, magnetic tape), an optical medium (such as a DVD), or a semiconductor medium. The semiconductor medium can be a solid-state drive.

[0149] It should be understood that in various embodiments of the present application, the magnitude of the sequence numbers of the above processes does not indicate the order of execution, and the order of execution of each process should be determined by its function and internal logic, and should not constitute any limitation to the implementation process of the embodiments of the present application.

[0150] Those of ordinary skill in the art can realize that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present application.

[0151] As described above, it is only the specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present application can easily think of changes or substitutions, which should all be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claimed rights.

Claims

1. A network optimization method based on online meetings, characterized in that, It includes the following steps: S1. In a cross-border network, deploy monitoring probes at the nodes participating in the online meeting to monitor network parameters such as network latency, bandwidth, packet loss rate, and jitter in real time. Use statistical analysis algorithms and time series models to process and predict anomalies in network parameters to obtain network condition information. Then, combined with the global network topology, use the reinforcement learning Q-Learning model and the graph neural network GNN topology prediction algorithm to output the optimal transmission path; S2. During the transmission of audio and video data along the optimal transmission path, use LSTM to predict bandwidth fluctuations, dynamically adjust the video resolution through the contextual bandit model, and adopt the improved weighted fair queueing WFQ and random early detection RED to preferentially transmit packets with high priorities, and predict the packet loss rate based on the Bayesian network, dynamically allocate redundant packets according to the frame type, and set the delay-constrained ARQ retransmission mechanism; S3. Use the PTP algorithm to process the device's local clock, fuse the clock signals of GPS / Beidou satellite timekeeping through Kalman filtering, correct the device clock deviation in real time, and use the adaptive synchronization buffer ASB and PID controller to dynamically adjust the audio and video playback timeline to synchronously control the audio and video stream and compensate for the decoding time difference; S4. Based on compliance labels, use the Kubernetes edge scheduler to dynamically deploy edge nodes. While ensuring that data is processed in compliance areas, after processing the audio and video data of the online meeting using the H.264 / H.265 video encoding and Opus audio encoding algorithms, use the ROI-driven compression and caching decision model combined with the international bandwidth cost to dynamically select the compression ratio and pre-cache meeting materials; S5. Collect user feedback data and metric data on the average latency, packet loss rate, and audio-visual synchronization error that affect the meeting performance. Use the deep deterministic policy gradient DDPG algorithm to dynamically adjust the routing and network parameters according to the real-time performance reward value of the meeting.

2. The network optimization method based on online meetings according to claim 1, wherein The steps for obtaining the network condition information are as follows: Deploy monitoring probes at n nodes in a transnational network to collect network parameters such as network delay, bandwidth, packet loss rate, and jitter in real time, which are calibrated as NP n ={D, B, L, J} n ; The statistical analysis algorithm calculates the mean and standard deviation of the network parameter NP, and uses the Z-score algorithm for anomaly detection. Among them, the calculation formula for the mean is and i ∈ n, where μ NP represents the mean of the network parameter NP, n represents the total number of nodes in the multinational network, i represents the i-th node in the total number of nodes, and NP i represents the network parameter value collected corresponding to the i-th node; The calculation formula for the standard deviation is In the formula, σ NP represents the standard deviation of the network parameter NP; The expression of the Z-score algorithm is When At this time It is expressed as the network parameter NP collected corresponding to the i-th node i Determine abnormality; Use the time series prediction ARIMA model to predict the bandwidth trend of the sliding window at the next moment. Among them, the calculation formula for the ARIMA model to predict the bandwidth trend is B t+1 = φ1B t + θ1∈ t + ∈ t+1 , where B t+1 represents the bandwidth value predicted by the ARIMA model at the t+1 moment, B t represents the bandwidth value of the time series of the bandwidth parameter at the current t moment, φ1 represents the autoregressive coefficient, θ1 represents the moving average coefficient, t represents the error term of white noise at the current t moment, t+1 represents the error term of white noise at the t+1 moment; Generate network condition information by extracting the characteristic information of bandwidth volatility, jitter accumulation value, and packet loss rate trend, and use the network condition information as the input of the subsequent model. Among them, the calculation formula of bandwidth volatility is and \(t\in T\). In the formula, \(T\) represents the size of the sliding window, \(\Delta B\) represents the bandwidth volatility, and \(B{1,2，…，T \}\) represents the bandwidth sequence within the time window \(T\), \(\max(B {1,2，…，T} )\), \(\min(B {1,2，…，T} )\), \(\mu(B {1,2，...，T} )\) respectively represent the maximum value, minimum value, and average value of the bandwidth sequence \(B {1,2，…，T} \) within the time window \(T\); The calculation formula for the jitter accumulation value is where J' represents the jitter accumulation value, J t represents the jitter value at time t, J t-1 represents the initial jitter value, and J t-1 = 0; The calculation formula for the packet loss rate trend is and In the formula, κ(L) represents the packet loss rate trend, represents the mean of the time indices within the time window T, μ L represents the mean of the packet loss rates within the time window T; The expression of the network condition information is NCI = (ΔB, J', κ(L)), where NCI represents the network condition information.

3. The network optimization method based on an online meeting according to claim 2, wherein The steps for outputting the optimal transmission path are as follows: Define the state space and action space in reinforcement learning according to the network parameters. Among them, the expression of the state space is Ss = {D, B, L, J, Cm}, where Ss represents the state space with the compliance label Cm inserted into the network parameters, and its dimension is 5×1. Cm represents the compliance label indicating whether the path violates data sovereignty, and the boolean value of Cm is 0 or 1; The expression of the action space is As = {switch the current path, keep the current path}, where As represents the action space; Design a multi-objective reward function and use the Q-Learning algorithm for iterative learning to update the Q-value table. Among them, the expression of the multi-objective reward function is R = -(αD + βL + γCm), and α + β + γ = 1, Cm = 0. Here, R represents the reward function, and α, β, γ represent the weight coefficients corresponding to the network delay D, packet loss rate L, and the non-compliance label Cm of violating data sovereignty; The calculation formula for updating the Q value is In the formula, Q′(Ss,As) represents the updated Q value, Q(Ss,As) represents the Q value of taking action As in state Ss, η represents the learning rate, and η ∈ (0, 1], λ represents the discount factor, represents the maximum Q value among all actions As in the next state Ss; Abstract the global network topology into a graph structure, where nodes represent devices such as routers and servers in the network, and edges represent the links connecting the devices. Then the attributes of each edge are network delay D and bandwidth B. Among them, the expression of the graph structure is Gs = (v i , e j ), where v i represents the nodes of network devices, and e j represents the attributes of the links; In the graph structure, according to the message passing mechanism of the audio-visual data in the GNN topology, the graph neural network GNN is used to predict the link congestion probability, and combined with the action selection of the reinforcement learning Q-Learning, the optimal transmission path is output. Among them, the calculation formula of the link congestion probability is Pc = σ(ω P ·h(e j ) + b P ), and Pc ∈ [0, 1]. In the formula, Pc represents the link congestion probability, σ represents the Sigmoid activation function, ω P represents the weight coefficient of the link congestion prediction layer, h(e j ) represents the embedding of the link attribute e j in the GNN message passing mechanism, and b P represents the bias term of the link congestion prediction layer; The expression for action selection in Q-Learning is where "As" represents the action selected for Q-Learning, and argmax() represents the argmax function for taking the maximum value; The output formula of the optimal transmission path is where p* represents the optimal transmission path, represents the set of all feasible paths in the graph structure, argmin() represents the argmin function for minimizing, D(e j ) represents the delay of the link, φ represents the indicator function for the link congestion probability Pc, and Pc > 0.7 represents high link congestion.

4. A network optimization method based on online meetings according to claim 3, characterized in that, The steps of adjusting the video resolution by the contextual slot machine model are as follows: Using a contextual bandit model, the bandwidth fluctuation is predicted from the network condition information NCI using LSTM. According to the current bandwidth, the current network jitter, and the priority of the defined frame type, the real-time network state features are extracted and labeled as nsc t = [ΔB t , B t , J t , FP t T , where nsc t represents the network state feature at the current time t, ΔB t represents the bandwidth change rate predicted by LSTM for a short period in the future, B t represents, J t represents, FP t represents the frame priority at the current time t, and T represents the transpose;​ Select the resolution to define the bitrate level Design the reward function for resolution to balance image quality and lag. Among them, the expression of the reward function for resolution is and ω1 + ω2 + ω3 = 1. In the formula,[[]] represents the reward function for resolution, ω1 represents the reward weight coefficient corresponding to the image quality,[[]] represents the image quality, ω2 represents the penalty weight coefficient corresponding to the lag,[[]] represents the lag, and the value of is 0 or 1, ω3 represents the penalty weight coefficient corresponding to the bandwidth matching,[[]] represents the bandwidth matching; Packets are preferentially queued and sent according to the frame priority FP using improved weighted fair queueing (WFQ) and random early detection (RED), and the packet loss threshold is dynamically adjusted. The ε-greedy strategy is used to select the bitrate level to determine the optimal transmission video resolution. The calculation formula for the optimal transmission video resolution is In the formula,[[]] represents the optimal transmission video resolution,[[]] represents the bitrate level[[]] Value function estimation.[[]] 5. The network optimization method based on online meetings according to claim 4, wherein The steps of the PTP algorithm processing device are as follows: In the network, determine the master clock and slave clock of the device. First, use the master clock to send a Sync message and record the master clock transmission timestamp as After receiving the synchronization Sync message from the clock, record the reception timestamp as The slave clock sends a Delay_Req message to the master clock and records the slave clock transmission timestamp as After the master clock receives the Delay_Req message, record the master clock reception timestamp as Send to the slave clock; The slave clock calculates the clock deviation and network latency from the master clock based on the received timestamp. Among them, the calculation formula for the clock deviation is In the formula, Offset represents the clock deviation of the slave clock relative to the master clock; The calculation formula for network latency is In the formula, Delay represents the network latency between the master and slave clocks; The slave clock adjusts its local clock according to the calculated clock deviation so that the slave clock is synchronized with the master clock.

6. The network optimization method based on online meetings according to claim 5, wherein The steps for correcting the device clock deviation are as follows: While collecting the clock signal of GPS / Beidou satellite time service, obtain the clock deviation and frequency deviation between the local clock and the GPS / Beidou satellite reference clock, define the clock state vector and calculate the clock state equation simultaneously. Among them, the expression of the clock state vector is In the formula, x k is expressed as the clock state vector, Δt k is expressed as the clock deviation, Δf k is expressed as the frequency deviation; The calculation formula for the clock state equation is where x k+1 represents the clock state at the next moment, represents the state transition matrix, and T′ represents the filtering period, w k represents the process noise that follows a Gaussian distribution; Based on the PTP algorithm, the time deviation of the local clock is obtained, and the time deviation measured by the GPS / Beidou satellite signal is fused to construct an observation equation. The Kalman filtering algorithm is used for iterative prediction and update to obtain the estimated value of the clock deviation of the posterior state. Among them, the expression of the observation equation is z k =[1, 0]x k +v k , and Δt k =Δt k PTP +Δt k GPS , where z k represents the observation equation, [1, 0] represents the observation matrix, v k represents the observation noise that follows a Gaussian distribution, Δt k PTP represents the time deviation of the local clock obtained by the PTP algorithm, and Δt k GPS represents the time deviation measured by the GPS / Beidou satellite signal; The local clock is adjusted for time and frequency in real time based on the clock deviation estimate of the a posteriori state.

7. The network optimization method based on online meetings according to claim 6, wherein, The processing steps of the ROI-driven compression and caching decision model are as follows: Construct a bandwidth cost model according to the dynamic changes of international bandwidth costs with links and time periods. Among them, the expression of the bandwidth cost model is In the formula, C BW represents the total bandwidth cost, Volume t represents the amount of transmitted data in time period t, Price t represents the bandwidth unit price in time period t, and T0 represents the total time period length; After processing the audio and video data of an online meeting using the H.264 / H.265 video encoding and Opus audio encoding algorithms, the resulting compressed data will reduce the picture quality. The peak signal-to-noise ratio (PSNR) is used to quantify the loss value of the picture quality. Among them, the calculation formula for the loss value of the picture quality is In the formula, ΔPQ represents the loss of picture quality after processing the audio and video data, MSE represents the mean square error of the image after compressing the audio and video data, 255 represents the maximum pixel value, and PSNR0 represents the PSNR value of the original audio and video; Calculate the return on investment ROI by combining the bandwidth cost and the loss value of the picture quality, and obtain the compression ratio decision rule by defining the constraint conditions according to the compression and caching decision model driven by ROI. Among them, the calculation formula of the return on investment ROI is and In the formula, ROI represents the return on investment, ΔC BW represents the bandwidth cost savings value, ζ represents the picture quality loss weight coefficient, ζ·ΔPQ represents the picture quality loss cost savings value, TC represents the total cost of bandwidth and picture quality, and CR represents the compression ratio; The expression of the compression ratio decision rule is In the formula, CR′ represents taking the maximum compression ratio; And a caching revenue target is formulated through an ROI-driven compression and caching decision model to obtain caching decision rules for pre-caching high-frequency conference materials to reduce duplicate transmission costs. Among them, the calculation formula for the caching decision rule is CC m = p m · s m · C BW , where CC m represents the expected revenue for caching the m-th conference material, p m represents the access probability of the m-th conference material, s m represents the size of the storage space occupied by the m-th conference material, and m represents the number of pre-cached conference materials.

8. A network optimization method based on an online meeting according to claim 7, characterized in that, The dynamic adjustment logic of the deep deterministic policy gradient DDPG algorithm is as follows: Collect user feedback data and metrics that affect conference performance, including average latency, packet loss rate, and audio-video synchronization error; Integrate the collected indicator data to define the status of the current meeting; Design a reward function to calculate the reward value based on the real-time performance status of the meeting; Using the policy network and value network in the DDPG algorithm, the policy network outputs an action based on the current state, and the policy network evaluates the value of taking an action under the state. The actions include route switching, FEC redundancy, and bit rate adjustment. Store each state, action, reward, and next state into the experience replay buffer; Randomly sample a batch of data from the experience replay buffer and define the loss function of the policy network to update the parameters of the policy network and value network; Based on the updated policy network, the optimal action is selected in the new state and applied to the conference system to dynamically adjust routing and network parameters; Iterate the DDPG algorithm steps to continuously collect data, calculate rewards, update network parameters, and adjust routing and network parameters.

9. A computer device, comprising: A memory and a processor; the memory stores a computer program, characterized in that when the processor executes the computer program, the steps of a network optimization method based on an online meeting according to any one of claims 1 to 8 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of a network optimization method based on an online meeting according to any one of claims 1 to 8 are implemented.

Images