Real-time audio and video transmission quality optimization method supporting large-scale concurrency

Abstract

The invention discloses a real-time audio and video transmission quality optimization method supporting large-scale concurrency, and relates to the technical field of computer networks, and the method comprises the specific steps: completing basic configuration after a terminal is started, constructing a bandwidth aggregation pool, collecting parameters, and initializing a neural network; inputting the link quality data into a network to generate a transmission strategy, and distributing FEC redundancy; matching an optimal node through a multi-stage scheduling tree; monitoring node state distribution tasks in real time; the terminal interacts data with the scheduling layer in real time, an end-network collaborative optimization closed loop is formed by adjusting strategies, and the audio and video transmission quality in a large-scale concurrent scene is guaranteed; the method comprises the following steps of: dynamically aggregating a network interface through a terminal, accurately pre-judging a link state, and constructing a self-adaptive transmission strategy system in combination with a dynamic FEC redundancy allocation and rapid switching mechanism; meanwhile, efficient resource utilization and accurate task allocation are realized by utilizing a multi-level scheduling architecture and a dynamic adaptation algorithm, million-level concurrent transmission is supported, and stability and service quality are improved.

Description

Technical Field

[0001] This invention relates to the field of computer network technology, specifically to a method for optimizing the quality of real-time audio and video transmission that supports large-scale concurrency. Background Technology

[0002] With the rapid development of the digital economy, real-time audio and video transmission technology has been widely applied to various fields such as remote collaborative office work, online education, e-commerce live streaming, video conferencing, and telemedicine. Users have placed higher demands on the real-time performance, stability, and clarity of audio and video transmission. At the same time, the service carrying capacity requirements in large-scale concurrent scenarios continue to grow. For example, scenarios such as cross-regional video conferencing for large enterprises, live classrooms with tens of thousands of participants on online education platforms, and live promotional events for e-commerce platforms all require supporting the simultaneous access of massive numbers of terminals and ensuring transmission quality. Currently, terminal devices generally have multiple types of network interfaces, and different network interfaces vary in terms of bandwidth and stability. Furthermore, network link status is susceptible to environmental interference, resulting in problems such as packet loss and latency fluctuations. Traditional transmission technologies are unable to fully integrate multi-interface resources and cope with dynamic changes in the link. How to achieve efficient real-time audio and video transmission in large-scale concurrent scenarios under complex network environments has become a key technical problem that urgently needs to be solved in the industry, and it has also driven the continuous exploration of related technologies such as transmission strategy optimization, node scheduling, and end-to-end network collaboration.

[0003] Traditional real-time audio and video transmission technologies often rely on a single network interface for data transmission, failing to effectively integrate the bandwidth resources of multiple network interfaces on the terminal. This results in low bandwidth utilization and difficulty in meeting the bandwidth demands of large-scale concurrent scenarios. Regarding link quality, they lack the ability to accurately predict link packet loss rates and other conditions. FEC redundancy configurations often use fixed schemes, unable to dynamically adjust based on link fluctuations and content importance, easily leading to insufficient redundancy causing the loss of critical content, or excessive redundancy consuming bandwidth resources. At the node scheduling level, the scheduling architecture is simple and struggles to efficiently break down millions of nodes. High-level concurrent requests can easily lead to uneven load distribution on edge nodes, with some nodes overloaded while others are idle. In terms of task allocation and fault handling, the task allocation logic lacks dynamic adaptability, failing to flexibly adjust the task type and quantity based on node load status. Furthermore, the link fault detection and switching response speed is slow, and data transmission interruption recovery takes a long time. In addition, there is a lack of efficient real-time data interaction mechanism between the terminal and the scheduling layer, making it impossible to form a closed loop for end-to-end collaborative optimization. This results in insufficient adaptability of the overall transmission system to dynamic changes, and in large-scale concurrent scenarios, problems such as audio and video stuttering and increased latency are likely to occur, affecting the user experience. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a method for optimizing the quality of real-time audio and video transmission that supports large-scale concurrency. This method automatically builds a dynamic bandwidth aggregation pool at the terminal, collects link parameters and initializes a lightweight temporal neural network, uses this network to predict the trend of link packet loss rate, dynamically allocates FEC redundancy, realizes node scheduling matching and dynamic task allocation through a distributed scheduling layer, and performs end-to-end network collaborative optimization based on real-time data interaction to form a closed-loop adjustment mechanism.

[0005] This method effectively improves the quality and stability of audio and video transmission in high-concurrency scenarios.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a method for optimizing real-time audio and video transmission quality that supports large-scale concurrency, the specific steps of which are as follows: S100, Basic Terminal Configuration: After the terminal starts up, it automatically identifies available network interfaces and builds a dynamic bandwidth aggregation pool, collects real-time link parameters, analyzes audio and video coding layer features and audio frame features, and performs initialization and parameter calibration of a lightweight temporal neural network. S200, Transmission strategy generation: The collected link quality data is input into a lightweight temporal neural network to calculate the future link packet loss rate trend. The FEC redundancy configuration scheme for each link with different priorities is generated through a dynamic asymmetric FEC redundancy allocation algorithm. The FEC redundancy configuration scheme includes the number of redundant packets or the redundancy ratio, and sets the link failure judgment conditions and fast switching mechanism. S300, node scheduling matching: Synchronizes link quality data, packet loss rate trend, link comprehensive value coefficient, FEC redundancy configuration scheme and its own geographical location information to the distributed scheduling layer. The distributed scheduling layer decomposes concurrent requests through a multi-level scheduling tree architecture and selects the optimal edge node through the node-task dynamic adaptation algorithm to complete the connection routing between the terminal and the node. S400, Dynamic Task Allocation: Monitors the hardware resource load status of each edge node in real time to the distributed scheduling layer, verifies the compatibility between terminal tasks and nodes based on the node-task dynamic adaptability algorithm, executes task allocation logic, and triggers a task migration mechanism when the node load exceeds the limit or the adaptability is insufficient. S500, End-to-Network Collaborative Optimization: The terminal and the distributed scheduling layer maintain real-time data interaction. The terminal continuously reports link quality fluctuations and FEC configuration effects, while the scheduling layer reports node load and processing status, and implements adjustments to link level, FEC redundancy configuration, node allocation, and task allocation strategies to form an end-to-network collaborative optimization closed loop.

[0007] Furthermore, in S100, the construction process of the dynamic bandwidth aggregation pool in the terminal's basic configuration is as follows: The terminal sequentially detects the physical connectivity of the 4G / 5G modem, WiFi 6 / 7 network card, and gigabit wired Ethernet adapter through the network interface scanning module, sends three connectivity probe packets via the ICMP protocol, and sets the timeout to 500ms. Interfaces with at least two successful probe responses, real-time bandwidth ≥1Mbps, and initial packet loss rate ≤10% are determined as valid interfaces. The MPTCP enhanced link binding protocol is used to bind all valid interfaces through sequence number synchronization and congestion control coordination mechanisms. Simultaneously, the available bandwidth of a single link is collected at a period of 50ms using the SNMP protocol. Total bandwidth of the aggregation pool ,in, Indicates the available bandwidth of a single link. This represents the total bandwidth of the aggregation pool.

[0008] Furthermore, in S100, the specific process of parsing the audio and video coding layer features and audio frame features in the terminal basic configuration is as follows: Connecting to the terminal's audio and video coding interface, reading the Slice partitioning identifier in the video bitstream to determine the Slice type; extracting GOP sequence information from the bitstream, counting the number of frames between consecutive keyframes to obtain the GOP period; identifying the identifier fields of intra-frame prediction mode and inter-frame prediction mode through pre-decoding bitstream analysis to clarify the corresponding prediction mode; using motion vector extraction technology to separate motion vector data from the video bitstream, and obtaining the motion vector amplitude through quantization processing; for audio frames, obtaining the sampling rate and bit rate parameters through audio coding protocol parsing; performing Fourier transform on the audio signal to analyze the frequency component distribution and energy proportion to determine the harmonic structure characteristics; and using speech activity detection technology to analyze the audio frame signal to determine whether the audio frame is in an active or silent state.

[0009] Furthermore, in S100, the specific process of deploying a lightweight temporal neural network and calibrating its parameters in the basic terminal configuration is as follows: a single hidden layer temporal neural network is selected as the packet loss rate prediction model, the hidden layer uses long short-term memory units or gated recurrent units, and the hidden layer has sixty-four neurons; pre-trained weights based on more than 100,000 multi-scenario link quality time-series data are loaded; thirty sets of real-time link latency, jitter, and packet loss rate data acquired within 300 milliseconds and a 10-millisecond acquisition cycle after the terminal starts are used as calibration samples; the gradient descent method is used to fine-tune the weights of the model output layer until the average absolute error between the predicted value and the actual sample value is reduced to 3% or less; a 50-millisecond sliding window is set to slice the real-time link data, and the packet loss rate trend for the next 200 to 500 milliseconds is output with each slide; the risk level is divided into low risk (≤3%), medium risk (3% to 8%), and high risk (>8%), and an identifier is output.

[0010] Furthermore, in S200, during the transmission strategy generation, the expression for the dynamic asymmetric FEC redundancy allocation algorithm is: ,in, It is FEC redundancy. This is the comprehensive value coefficient of the link. This is the content time sensitivity coefficient. For redundant baseline coefficients, This is the predicted packet loss rate for the link. As a feedback adjustment factor, This represents the percentage of key content lost in the previous period. For edge node computing power redundancy, This is the threshold for node computing power redundancy. It is the floor operator.

[0011] Furthermore, in S200, the specific process of the link failure switching mechanism during transmission strategy generation is as follows: the terminal monitors the real-time packet loss rate of each link in real time. With real-time latency When a certain link satisfies or If a link failure is detected, the link with the highest comprehensive value coefficient among the currently available links is immediately retrieved as the target switching link. The audio and video data carried by the failed link is migrated to the target link using sequence number alignment technology, while maintaining the continuity of data transmission sequence numbers during the migration process.

[0012] Furthermore, in S300, during node scheduling and matching, the specific process of the distributed scheduling layer decomposing concurrent requests through a multi-level scheduling tree architecture is as follows: The multi-level scheduling tree architecture consists of a root node, regional nodes, and edge nodes. The root node receives concurrent requests sent by terminals across the entire network, classifies and statistically analyzes the requests, determines the geographical region to which each request belongs based on the geographical location information synchronized by the terminal, and distributes the concurrent requests of the corresponding region to the regional nodes corresponding to that geographical region in batches. The regional nodes receive the concurrent requests distributed by the root node within their respective regions, collect the load status data of all edge nodes within their jurisdiction in real time, including CPU utilization, memory usage, and the number of current connections, sort the load status of each edge node, filter out edge nodes with low load ranges, and evenly distribute the concurrent requests within their respective regions to the filtered edge nodes according to the processing capacity of the edge nodes, thereby realizing the decomposition and allocation of millions of concurrent requests to thousands of edge nodes.

[0013] Furthermore, in S300, during node scheduling and matching, the expression for the node-task dynamic fit algorithm is: ,in, It is the dynamic adaptability of nodes and tasks. This is the comprehensive value coefficient of the link. For FEC redundancy, This represents the node task processing efficiency coefficient. To adapt to the weighting coefficients, To standardize backend-to-end latency, This refers to the actual end-to-end latency from the node to the terminal. This represents the total complexity of the terminal task. To determine the remaining processing capacity of a node, a fitness threshold of 70% is set, and nodes with fitness values ​​not lower than this threshold are selected as target candidate nodes. From the target candidate nodes, the node with the lowest end-to-end latency, CPU utilization not exceeding 60%, and memory usage not exceeding 70% is selected as the optimal edge node.

[0014] Furthermore, in S400, the specific process of executing task allocation logic according to node load intervals in dynamic task allocation is as follows: Node computing power redundancy is calculated based on node hardware resource load status data, and three load intervals are divided based on CPU utilization and memory utilization. The first load interval is when CPU utilization is below 50% and memory utilization is below 60%. In this interval, nodes are in a low-load state, and combined tasks of encoding, mixing, and FEC redundancy calculations are allocated. The number of combined tasks simultaneously carried by a single node does not exceed 500. The second load interval is when CPU utilization is above 50%. Between 60% and 80% CPU utilization or memory utilization, nodes are in a medium load state and are only assigned a single mixing task. The number of mixing tasks that a single node can simultaneously handle does not exceed 1000. The third load range is when CPU utilization is above 80% or memory utilization is above 80%. In this range, nodes are in a high load state, and the allocation of new tasks is suspended. Only the tasks that have been assigned but not yet completed are processed. During the task allocation process, the total complexity of the terminal tasks and the remaining processing capacity of the node are considered to ensure that the total computing power consumption of all types of tasks on a single node does not exceed 90% of the node's remaining processing capacity.

[0015] Furthermore, in the S500, the specific mechanism for real-time data interaction in the end-to-end network collaborative optimization is as follows: the terminal and the distributed scheduling layer use a TLS1.3 encrypted transmission channel, and the data transmission uses AES-256-GCM encryption. A fixed interaction period of 20ms is used. When the link packet loss rate changes by ≥2%, the latency fluctuation is ≥30ms, or the node CPU utilization is >90%, an emergency feedback is triggered. During emergency feedback, the interaction period is shortened to 10ms until the state returns to stability. After receiving the data, the scheduling layer uses CRC32 verification to ensure data integrity. If the verification fails, a retransmission mechanism is triggered, and the retransmission timeout is ≤10ms.

[0016] Compared with existing technologies, this method for optimizing real-time audio and video transmission quality that supports large-scale concurrency has the following advantages: I. This invention utilizes dynamic aggregation of network interfaces and collection of link parameters during the terminal's basic configuration phase. Combined with a lightweight temporal neural network for accurate prediction of link status, and a dynamic asymmetric FEC redundancy allocation mechanism and fast link switching logic, it constructs an adaptive transmission strategy system. By integrating multiple network interfaces to form bandwidth aggregation capabilities, it fully exploits available network resources. Simultaneously, based on link quality trend analysis and content feature adaptation, it adjusts redundancy configurations in a targeted manner. This ensures the integrity of critical content transmission while avoiding resource waste. In the event of link anomalies, it quickly restores data transmission through an efficient switching mechanism, effectively resisting the impact of link fluctuations on audio and video transmission. It significantly improves the stability and anti-interference capability of transmission in large-scale concurrent scenarios, ensuring continuous output of audio and video streams and improving user experience.

[0017] Second, this invention achieves efficient decomposition and optimal node matching of concurrent requests through a multi-level architecture design of the distributed scheduling layer and a node-task dynamic adaptation algorithm. Combined with dynamic task allocation and end-to-end network collaborative optimization closed loop, it maximizes resource utilization and concurrent processing efficiency. The hierarchical scheduling architecture rationally distributes massive concurrent requests to edge nodes, avoiding single-point overload. At the same time, it accurately allocates task types and quantities based on node load status and task adaptability, realizing dynamic matching of resources and needs. The optimized mechanism of real-time data interaction between the terminal and the scheduling layer can respond promptly to changes in link quality and node status fluctuations, dynamically adjust transmission and scheduling strategies, and form a full-process adaptive optimization, supporting efficient transmission in scenarios with millions of concurrent requests, balancing transmission efficiency and service quality, and expanding the applicable scenarios and carrying capacity of real-time audio and video applications.

[0018] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0020] Figure 1 A flowchart illustrating a method for optimizing the quality of real-time audio and video transmission that supports large-scale concurrency. Figure 2 This is a schematic diagram illustrating the data transmission steps of a method for optimizing the quality of real-time audio and video transmission that supports large-scale concurrency. Detailed Implementation

[0021] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0022] Example 1: Multi-hospital remote medical consultation scenario.

[0023] S100, Basic Terminal Configuration: A top-tier hospital, in collaboration with two branch hospitals, conducted remote medical consultations. The consultation terminals were dedicated high-definition medical audio and video equipment deployed in consultation rooms across the various hospital campuses. The network environment included gigabit wired Ethernet and a dedicated 5G network. Upon startup, the terminals automatically identified available network interfaces and sent three connectivity probe packets via the ICMP protocol with a timeout of 500ms. The gigabit wired Ethernet interface and the 5G network interface with at least two successful probe responses, real-time bandwidth ≥1Mbps, and initial packet loss rate ≤10% were selected as valid interfaces. The MPTCP enhanced link bonding protocol was used to bind the two valid interfaces through sequence number synchronization and congestion control mechanisms, constructing a dynamic bandwidth aggregation pool. This provided dual-link assurance for high-definition medical audio and video data transmission, reducing the impact of a single link failure on the consultation. Simultaneously, the available bandwidth of a single link and the total bandwidth of the aggregation pool were collected every 50ms using the SNMP protocol to monitor bandwidth changes in real time, ensuring that the bandwidth met the transmission requirements of large-capacity medical data such as lesion images.

[0024] The terminal interfaces with audio and video encoding, reads the slice segmentation identifier from the video stream to determine the slice type, extracts GOP sequence information, and counts the number of frames between consecutive keyframes to obtain the GOP period. Before decoding, the bitstream analysis identifies the identifier fields of intra-frame and inter-frame prediction modes to clarify the corresponding prediction modes. Motion vector extraction technology is used to separate motion vector data from the video stream and quantize it to obtain the motion vector amplitude. Emphasis is placed on extracting clarity-related features of the lesion area video, laying the foundation for accurate subsequent transmission of lesion images. For audio frames, the sampling rate and bit rate parameters are obtained through audio encoding protocol parsing. Fourier transform analysis is performed on the audio signal to determine the frequency component distribution and energy ratio to identify harmonic structure characteristics. Speech activity detection technology is used to determine whether the audio frame is in an active or silent state, accurately distinguishing key audio content in medical communication and ensuring priority and clear transmission of diagnostic communication between doctors.

[0025] A single hidden-layer temporal neural network is deployed on the terminal as the packet loss rate prediction model. The hidden layer uses gated recurrent units with 64 neurons and is loaded with pre-trained weights trained on over 100,000 multi-scenario link quality time-series data, enabling the model to predict link states adaptably to medical scenarios. Thirty sets of real-time link latency, jitter, and packet loss rate data acquired within 300 milliseconds of terminal startup and collected at 10-millisecond intervals are used as calibration samples. Gradient descent is used to fine-tune the model's output layer weights until the average absolute error between the predicted value and the actual sample value is reduced to 3% or less, improving the model's prediction accuracy in the current consultation network environment. A 50-millisecond sliding window slices real-time link data; each slide outputs the packet loss rate trend for the next 200 to 500 milliseconds. Packet loss rates are categorized as low risk (≤3%), medium risk (3% to 8%), and high risk (>8%), with corresponding labels. This provides early risk warnings for medical data transmission, facilitating timely adjustments to transmission strategies to protect medical data. Figure 1 As shown.

[0026] S200, Transmission Strategy Generation: The terminal inputs the collected link quality data into a lightweight temporal neural network to accurately calculate future link packet loss rate trends, providing a reliable basis for medical data transmission strategy formulation. Considering the high timeliness and importance of medical audio and video data, a dynamic asymmetric FEC redundancy allocation algorithm is used to generate FEC redundancy configuration schemes for different priority content on each link. The expression for the dynamic asymmetric FEC redundancy allocation algorithm is: ,in, It is FEC redundancy. This is the comprehensive value coefficient of the link. This is the content time sensitivity coefficient. For redundant baseline coefficients, This is the predicted packet loss rate for the link. As a feedback adjustment factor, This represents the percentage of key content lost in the previous period. For edge node computing power redundancy, This is the threshold for node computing power redundancy. It uses an up-rounding operator to specify the number or ratio of redundant packets. High redundancy is configured for core medical data such as high-resolution lesion images and key diagnostic statements from doctors, ensuring that even if packets are lost during transmission, the data can be recovered through redundant packets. Reasonable redundancy is configured for non-core data such as routine environmental images to avoid excessive bandwidth consumption affecting core data transmission. Simultaneously, link failure judgment conditions and a rapid switching mechanism are set. The terminal monitors the real-time packet loss rate and real-time latency of each link. When a link meets the criteria of a real-time packet loss rate > 15% or a real-time latency > 300ms, it is judged as a link failure. The link with the highest comprehensive value coefficient among the currently available links is immediately selected as the target switching link. Medical audio and video data carried by the failed link is migrated to the target link using sequence number alignment technology. During the migration process, the data transmission sequence number is maintained continuously, and the total time from failure detection to data recovery is strictly controlled to within 50ms, avoiding interruptions to the consultation process due to link failures and ensuring the continuity of the diagnostic process.

[0027] S300, node scheduling matching: The terminal synchronizes link quality data, packet loss rate trends, link comprehensive value coefficient, FEC redundancy configuration scheme, and its own geographical location information to the distributed scheduling layer, providing comprehensive data support for the scheduling layer to accurately match nodes. The distributed scheduling layer decomposes concurrent requests through a multi-level scheduling tree architecture. The root node receives concurrent requests from terminals in the three hospital areas, classifies and statistically analyzes them, determines the geographical region to which each request belongs based on the geographical location information synchronized by the terminal, and distributes concurrent requests from the corresponding region to the corresponding regional nodes in batches, reducing latency caused by cross-regional transmission and allowing medical data to quickly reach the regional nodes. Regional nodes receive concurrent requests distributed by the root node within their region, collect real-time load status data such as CPU utilization, memory utilization, and current number of connections of all edge nodes within their jurisdiction, sort the load status of each edge node, and filter out edge nodes with low load. Concurrent requests within their region are then evenly distributed to the selected edge nodes according to their processing capacity, avoiding uneven load distribution that could affect processing efficiency. The optimal edge node is selected using a node-task dynamic adaptation algorithm, the expression of which is: ,in, It is the dynamic adaptability of nodes and tasks. This is the comprehensive value coefficient of the link. For FEC redundancy, This represents the node task processing efficiency coefficient. To adapt to the weighting coefficients, To standardize backend-to-end latency, This refers to the actual end-to-end latency from the node to the terminal. This represents the total complexity of the terminal task. To determine the remaining processing capacity of nodes, a fitness threshold of 70% is set, and nodes with fitness values ​​not lower than this threshold are selected as target candidate nodes. From the target candidate nodes, the node with the lowest end-to-end latency, CPU utilization not exceeding 60%, and memory utilization not exceeding 70% is selected as the optimal edge node. The fitness value of each candidate node is calculated, and target candidate nodes with fitness values ​​not lower than 70% are selected. Then, from the target candidate nodes, the node with the lowest end-to-end latency, CPU utilization not exceeding 60%, and memory utilization not exceeding 70% is selected to complete the connection routing between the terminal and the node, ensuring that medical data can be transmitted to the node for processing with the lowest latency, providing low-latency support for real-time remote consultation.

[0028] S400, dynamic task allocation: The distributed scheduling layer monitors the hardware resource load status of each edge node in real time, and verifies the compatibility between the terminal consultation task and the node based on the node-task dynamic adaptation algorithm, ensuring that the consultation task is assigned to the node with the corresponding processing capability, and executes the task allocation logic according to the node load range. The redundancy of node computing power is calculated based on the node hardware resource load status data. Three load ranges are defined by combining CPU utilization and memory utilization: The first load range is when CPU utilization is below 50% and memory utilization is below 60%. In this range, nodes are under low load, and combined tasks such as encoding, mixing, and FEC redundancy calculations are assigned. The number of combined tasks simultaneously handled by a single node does not exceed 500, fully utilizing idle node resources to process complex consultation tasks. The second load range is when CPU utilization is between 50% and 80% or memory utilization is between 60% and 80%. In this range, nodes are under medium load, and only a single mixing task is assigned. The number of mixing tasks simultaneously handled by a single node does not exceed 1000, avoiding excessive node load and processing delays. The third load range is when CPU utilization is above 80% or memory utilization is above 80%. In this range, nodes are under high load, pausing the assignment of new tasks and only processing existing unfinished tasks to ensure stable processing of current consultation tasks. During task allocation, the total complexity of the consultation task is combined with the remaining processing capacity of the node to ensure that the total computing power consumption of various tasks on a single node does not exceed 90% of the node's remaining processing capacity, thereby further ensuring the stability of the node in handling consultation tasks. When the node load exceeds the limit or the adaptability is insufficient, the task migration mechanism is triggered to transfer the consultation task to a more suitable node, so as to avoid task interruption affecting the doctor's diagnosis.

[0029] S500, end-to-end network coordination optimization: The terminal maintains real-time data interaction with the distributed scheduling layer, employing a TLS 1.3 encrypted transmission channel and AES-256-GCM encryption for data transmission. This effectively protects patient privacy data from leakage or tampering during transmission, complying with medical data security standards. A fixed interaction cycle of 20ms ensures timely synchronization of link quality, FEC configuration effectiveness, node load, and other status information between the terminal and the scheduling layer. The terminal continuously reports link quality fluctuations and FEC configuration effectiveness, while the scheduling layer reports node load and processing status. Emergency feedback is triggered when the link packet loss rate changes by ≥2%, latency fluctuations by ≥30ms, or node CPU utilization exceeds 90%. During emergency feedback, the interaction cycle is shortened to 10ms until the status stabilizes, enabling rapid response to transmission or node anomalies and preventing the escalation of anomalies from impacting consultations. After receiving data, the scheduling layer uses CRC32 verification to ensure data integrity, preventing policy adjustment errors due to data corruption. If verification fails, a retransmission mechanism is triggered, with a retransmission timeout of ≤10ms to ensure timely retransmission of critical status data. Based on real-time adjustments to link levels, FEC redundancy configurations, node allocation, and task allocation strategies using interactive data, a closed-loop optimization mechanism is formed to ensure clear and smooth audio and video, stable and reliable data transmission during multi-hospital remote medical consultations, providing strong technical support for doctors' accurate diagnoses.

[0030] In summary, in multi-hospital telemedicine consultation scenarios, the terminal configuration phase involves building a dual-link dynamic bandwidth aggregation pool, accurately extracting medical audio and video features, optimizing the packet loss rate prediction model, employing dynamic asymmetric FEC for high redundancy of core medical data configuration, achieving low-latency node matching through a multi-level scheduling tree and node adaptation algorithm, dynamically allocating consultation tasks according to load, and ensuring data security with TLS 1.3 and AES-256-GCM. Figure 2 As shown, this method ensures clear and stable medical audio and video transmission, avoids interruptions in consultations, protects patient privacy, and provides reliable technical support for real-time communication and accurate diagnosis among doctors across hospital areas, meeting the high requirements of medical scenarios for transmission security and stability.

[0031] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for optimizing real-time audio and video transmission quality to support large-scale concurrency, characterized in that, The specific steps of this method are as follows: S100, Basic Terminal Configuration: After the terminal starts up, it automatically identifies available network interfaces and builds a dynamic bandwidth aggregation pool, collects real-time link parameters, analyzes audio and video coding layer features and audio frame features, and performs initialization and parameter calibration of a lightweight temporal neural network. S200, Transmission strategy generation: The collected link quality data is input into a lightweight temporal neural network to calculate the future link packet loss rate trend. The FEC redundancy configuration scheme for each link with different priorities is generated through a dynamic asymmetric FEC redundancy allocation algorithm. The FEC redundancy configuration scheme includes the number of redundant packets or the redundancy ratio, and sets the link failure judgment conditions and fast switching mechanism. S300, node scheduling matching: Synchronizes link quality data, packet loss rate trend, link comprehensive value coefficient, FEC redundancy configuration scheme and its own geographical location information to the distributed scheduling layer. The distributed scheduling layer decomposes concurrent requests through a multi-level scheduling tree architecture and selects the optimal edge node through the node-task dynamic adaptation algorithm to complete the connection routing between the terminal and the node. S400, Dynamic Task Allocation: Monitors the hardware resource load status of each edge node in real time to the distributed scheduling layer, verifies the compatibility between terminal tasks and nodes based on the node-task dynamic adaptability algorithm, executes task allocation logic, and triggers a task migration mechanism when the node load exceeds the limit or the adaptability is insufficient. S500, End-to-Network Collaborative Optimization: The terminal and the distributed scheduling layer maintain real-time data interaction. The terminal continuously reports link quality fluctuations and FEC configuration effects, while the scheduling layer reports node load and processing status, and implements adjustments to link level, FEC redundancy configuration, node allocation, and task allocation strategies to form an end-to-network collaborative optimization closed loop.

2. The method for optimizing real-time audio and video transmission quality supporting large-scale concurrency as described in claim 1, characterized in that, In S100, the construction process of the dynamic bandwidth aggregation pool in the terminal basic configuration is as follows: The terminal sequentially detects the physical connectivity of the 4G / 5G modem, WiFi 6 / 7 network card, and gigabit wired Ethernet adapter through the network interface scanning module. It sends three connectivity probe packets via the ICMP protocol with a timeout of 500ms. Interfaces with at least two successful probe responses, real-time bandwidth ≥1Mbps, and initial packet loss rate ≤10% are determined as valid interfaces. The MPTCP enhanced link binding protocol is used to bind all valid interfaces through sequence number synchronization and congestion control coordination mechanisms. The available bandwidth of a single link is collected synchronously via the SNMP protocol at a period of 50ms. Total bandwidth of the aggregation pool ,in, Indicates the available bandwidth of a single link. This represents the total bandwidth of the aggregation pool.

3. The method for optimizing real-time audio and video transmission quality supporting large-scale concurrency as described in claim 1, characterized in that, In S100, the specific process of parsing the audio and video coding layer features and audio frame features in the terminal basic configuration is as follows: Interact with the terminal's audio and video coding interface, read the Slice partitioning identifier in the video bitstream to determine the Slice type; extract GOP sequence information from the bitstream, count the number of frames between consecutive keyframes to obtain the GOP period; identify the identifier fields of intra-frame prediction mode and inter-frame prediction mode through pre-decoding bitstream analysis to clarify the corresponding prediction mode; use motion vector extraction technology to separate motion vector data from the video bitstream, and obtain the motion vector amplitude through quantization processing; for audio frames, obtain the sampling rate and bit rate parameters through audio coding protocol parsing; perform Fourier transform on the audio signal, analyze the frequency component distribution and energy ratio to determine the harmonic structure characteristics; use speech activity detection technology to analyze the audio frame signal and determine whether the audio frame is in an active or silent state.

4. The method for optimizing real-time audio and video transmission quality supporting large-scale concurrency as described in claim 1, characterized in that, In the S100 terminal basic configuration, the specific process of deploying a lightweight temporal neural network and performing parameter calibration is as follows: a single hidden layer temporal neural network is selected as the packet loss rate prediction model, the hidden layer adopts long short-term memory units or gated recurrent units, and the hidden layer has sixty-four neurons; pre-trained weights based on more than 100,000 multi-scenario link quality time-series data are loaded; thirty sets of real-time link latency, jitter, and packet loss rate data acquired within 300 milliseconds and a 10-millisecond acquisition cycle after the terminal starts are used as calibration samples; the gradient descent method is used to fine-tune the model output layer weights until the average absolute error between the predicted value and the actual sample value is reduced to 3% or less. Set a 50-millisecond sliding window slice for real-time link data, and output the packet loss rate trend for the next 200 to 500 milliseconds with each slide; The risk levels are classified as low risk (≤3%), medium risk (3% to 8%), and high risk (greater than 8%), and an identifier is output accordingly.

5. The method for optimizing real-time audio and video transmission quality supporting large-scale concurrency as described in claim 1, characterized in that, In S200, during the transmission strategy generation, the expression for the dynamic asymmetric FEC redundancy allocation algorithm is: ,in, It is FEC redundancy. This is the comprehensive value coefficient of the link. This is the content time sensitivity coefficient. For redundant baseline coefficients, This is the predicted packet loss rate for the link. As a feedback adjustment factor, This represents the percentage of key content lost in the previous period. For edge node computing power redundancy, This is the threshold for node computing power redundancy. It is the floor operator.

6. The method for optimizing real-time audio and video transmission quality supporting large-scale concurrency as described in claim 1, characterized in that, In step S200, the specific process of the link failure switching mechanism during transmission strategy generation is as follows: the terminal monitors the real-time packet loss rate of each link. With real-time latency When a certain link satisfies or When this happens, it is determined to be a link failure; Immediately retrieve the link with the highest comprehensive value coefficient among the currently available links and use it as the target switching link. Then, use sequence number alignment technology to migrate the audio and video data carried by the faulty link to the target link, maintaining the continuity of data transmission sequence numbers during the migration process.

7. The method for optimizing real-time audio and video transmission quality supporting large-scale concurrency as described in claim 1, characterized in that, In the S300 node scheduling matching process, the specific process of the distributed scheduling layer decomposing concurrent requests through the multi-level scheduling tree architecture is as follows: The multi-level scheduling tree architecture consists of a root node, regional nodes, and edge nodes. The root node receives concurrent requests sent by terminals across the network, classifies and counts the requests, determines the geographical region to which each request belongs based on the geographical location information synchronized by the terminal, and distributes the concurrent requests of the corresponding region to the regional nodes corresponding to that geographical region in batches. Regional nodes receive concurrent requests distributed by the root node within their region, collect load status data of all edge nodes within their jurisdiction in real time, including CPU utilization, memory usage, and current number of connections, sort the load status of each edge node, filter out edge nodes with low load, and evenly distribute concurrent requests within their region to the filtered edge nodes according to the processing capacity of each edge node, thereby achieving the distribution of millions of concurrent requests to thousands of edge nodes.

8. The method for optimizing real-time audio and video transmission quality supporting large-scale concurrency according to claim 1, characterized in that, In S300, during node scheduling and matching, the expression for the node-task dynamic fit algorithm is: ,in, It is the dynamic adaptability of nodes and tasks. This is the comprehensive value coefficient of the link. For FEC redundancy, This represents the node task processing efficiency coefficient. To adapt to the weighting coefficients, To standardize backend-to-end latency, This refers to the actual end-to-end latency from the node to the terminal. This represents the total complexity of the terminal task. To determine the remaining processing capacity of a node, a fitness threshold of 70% is set, and nodes with fitness values ​​not lower than this threshold are selected as target candidate nodes. From the target candidate nodes, the node with the lowest end-to-end latency, CPU utilization not exceeding 60%, and memory usage not exceeding 70% is selected as the optimal edge node.

9. The method for optimizing real-time audio and video transmission quality supporting large-scale concurrency according to claim 1, characterized in that, In the S400 task dynamic allocation, the specific process of executing task allocation logic according to node load intervals is as follows: The node computing power redundancy is calculated based on the node hardware resource load status data, and three load intervals are divided based on CPU utilization and memory utilization. The first load interval is when CPU utilization is below 50% and memory utilization is below 60%. In this interval, the node is in a low-load state, and a combination of encoding, mixing, and FEC redundancy calculation tasks is allocated. The number of combined tasks simultaneously carried by a single node does not exceed 500. The second load interval is when CPU utilization is between 50% and 80% or memory utilization is between 60% and 80%. In this interval, the node is in a medium-load state, and only a single mixing task is allocated. The number of mixing tasks simultaneously carried by a single node does not exceed 1000. The third load interval is when CPU utilization is above 80% or memory utilization is above 80%. In this interval, the node is in a high-load state, and the allocation of new tasks is paused, with only the already accepted but unfinished tasks being processed. During the task allocation process, the total complexity of the terminal tasks and the remaining processing capacity of the node are combined to ensure that the total computing power consumption of various tasks on a single node does not exceed 90% of the node's remaining processing capacity.

10. A method for optimizing real-time audio and video transmission quality supporting large-scale concurrency as described in claim 1, characterized in that, In the S500, the specific mechanism for real-time data interaction in the end-to-end network collaborative optimization is as follows: the terminal and the distributed scheduling layer use a TLS1.3 encrypted transmission channel, and the data transmission uses AES-256-GCM encryption. A fixed interaction period of 20ms is used. When the link packet loss rate changes by ≥2%, the latency fluctuation is ≥30ms, or the node CPU utilization is >90%, an emergency feedback is triggered. During emergency feedback, the interaction period is shortened to 10ms until the state returns to stability. After receiving the data, the scheduling layer uses CRC32 verification to ensure data integrity. If the verification fails, a retransmission mechanism is triggered, and the retransmission timeout is ≤10ms.

Images