Cloud conference audio and video resource dynamic allocation optimization method empowered by edge computing

Abstract

The present application relates to the technical field of computer communication and edge computing, and discloses a cloud conference audio and video resource dynamic allocation optimization method empowered by edge computing, which comprises the following steps: collecting multi-dimensional perception interaction signals and node fluctuation signals for separate processing to obtain a service flow priority state vector and a multi-dimensional resource state vector; collecting end-to-end timestamp alignment signals and historical bandwidth sequence data for separate processing to obtain a synchronization error scalar signal and a predicted bandwidth boundary vector; splicing and fusing to generate a system environment state joint matrix, and performing nonlinear penalty mapping on the synchronization error scalar signal to obtain a dynamic synchronization constraint factor; inputting the joint matrix into a double-delay deep deterministic policy gradient network, and combining the predicted bandwidth boundary vector and the dynamic synchronization constraint factor to correct a reward function. The present application realizes closed-loop feedback of multi-dimensional physical feature cooperation, reduces invalid computing power consumption, and improves transmission certainty of cloud conference under edge network.

Description

Technical Field

[0001] This invention relates to the field of computer communication and edge computing, specifically to a method for dynamic allocation and optimization of cloud conferencing audio and video resources enabled by edge computing. Background Technology

[0002] The collaborative architecture of cloud computing and edge computing pushes underlying physical computing resources down to the edge of the backbone network, shortening the spatial distance for multimedia data transmission at the physical level. Cloud conferencing systems involve real-time interaction of multiple high-concurrency real-time audio transmission units, video reference frames, and underlying control signaling. Heterogeneous multimodal data imposes specific physical constraints on transmission physical latency, image quality and frame rate stability, and the synchronization of multi-terminal signaling interaction.

[0003] Existing edge computing resource allocation architectures typically rely on threshold scheduling models based on static rules, or use only a single-dimensional network channel congestion feedback mechanism for coarse-grained resource allocation. When multiple physical terminal users frequently send commands to seize the right to speak during a cloud conferencing process, when the available physical bandwidth of the network backbone fluctuates tidal characteristics, or when the computing power of the general-purpose graphics processors of local edge base station computing nodes is at a critical state of high temperature and frequency reduction saturation, the single-dimensional scheduling mechanism cannot find an analytical solution between physical load balancing of computing nodes and strict time-domain alignment of audio and video timestamps.

[0004] Traditional buffering strategies forcibly increase memory stack latency at the network receiver to eliminate jitter, causing delays in terminal interaction signaling responses. Conventional reinforcement learning-based task offloading algorithms only focus on the overall throughput limit of the communication channel, failing to extract attention priorities that conform to the interaction logic from multi-party concurrent audio and video transmission streams. This results in the physical computing power of edge nodes being wasted on processing silent audio streams or inactive black-screen video streams. The current underlying control architecture lacks a physical closed-loop feedback system for joint matrix modeling of underlying time base synchronization errors, mid-layer physical channel bandwidth variation trends, and high-layer application programming interface conference control logic. Summary of the Invention

[0005] The purpose of this invention is to provide a dynamic allocation optimization method for cloud conferencing audio and video resources enabled by edge computing, so as to solve the technical problems of resource allocation instruction delay and ineffective consumption of computing power of physical edge nodes caused by the lack of comprehensive consideration of multi-dimensional heterogeneous physical characteristics and the lack of multi-level model physical collaborative closed loop in the existing technical architecture.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0007] The edge computing-enabled cloud conferencing audio and video resource dynamic allocation optimization method includes the following steps:

[0008] Step 1: Collect multi-dimensional perception interaction signals from the cloud conferencing platform, perform time-decay attention priority evaluation processing on the multi-dimensional perception interaction signals to obtain a business flow priority state vector; simultaneously collect node fluctuation signals from the edge computing physical environment, perform multi-dimensional feature extraction processing on the node fluctuation signals to obtain a multi-dimensional resource state vector.

[0009] Step 2: Acquire the end-to-end timestamp alignment signal of the physical receiving port, perform Kalman filtering clock drift compensation processing on the end-to-end timestamp alignment signal to obtain the synchronization error scalar signal; synchronously acquire the historical bandwidth sequence data of the underlying network link, perform bidirectional long short-term memory network feature extraction processing on the historical bandwidth sequence data to obtain the predicted bandwidth boundary vector;

[0010] Step 3: Perform feature matrix concatenation and alignment processing on the service flow priority state vector and the multidimensional resource state vector to obtain the system environment state joint matrix; perform nonlinear penalty mapping scaling processing on the synchronization error scalar signal to obtain the dynamic synchronization constraint factor.

[0011] Step 4: Input the joint matrix of system environment state into the dual-delay deep deterministic policy gradient network for global policy inference processing. Combine the predicted bandwidth boundary vector and the dynamic synchronization constraint factor to perform reward function correction interaction calculation on the evaluation network embedded in the dual-delay deep deterministic policy gradient network. Output resource allocation action vector for controlling physical edge nodes. Adjust the hardware computing power slice boundary of computing nodes according to the resource allocation action vector and dynamically refresh the hardware encoding quantization parameters of video stream, and trigger the termination judgment logic based on norm geometric distance.

[0012] As a preferred embodiment of the present invention, the step of performing time-decay attention priority evaluation processing on the multi-dimensional perceptual interaction signal to obtain the business flow priority state vector includes the following specific processing logic:

[0013] Extract the physical time interval variable and logical Boolean indicator variable for concurrent users;

[0014] The physical time interval variable is substituted into the preset attention time decay rate control coefficient to perform an exponential decay operation, resulting in an exponential decay term;

[0015] The logical Boolean indicator variable is combined with the corresponding system gain correction coefficient and an accumulation operation is performed to obtain the gain accumulation term;

[0016] By integrating the network role basic permission weight coefficient, the exponential decay term, and the gain accumulation term, a priority status evaluation value is obtained; and a one-dimensional floating-point array format of the service flow priority status vector is generated by traversing all concurrent access physical terminals.

[0017] As a preferred embodiment of the present invention, the multi-dimensional feature extraction processing of the node fluctuation signal to obtain a multi-dimensional resource state vector includes the following specific processing logic:

[0018] Extract the true transient sample values, lower limit reference constant, and upper limit threshold constant of various node physical characteristics from the node fluctuation signal;

[0019] The difference between the actual transient sample value and the lower limit reference constant is calculated as the numerator;

[0020] The difference between the upper limit threshold constant and the lower limit benchmark constant is calculated as the denominator;

[0021] Divide the numerator by the denominator to obtain a dimensionless mapping value;

[0022] The multidimensional resource state vector is generated by combining the dimensionless mapping values ​​of all collected dimensions.

[0023] Furthermore, the end-to-end timestamp alignment signal is subjected to Kalman filtering clock drift compensation processing to obtain a synchronization error scalar signal. The specific processing logic includes:

[0024] Extract the prior physical error covariance matrix, the observation mapping mathematical matrix, and the measurement noise covariance matrix, and calculate the Kalman filter gain matrix variables;

[0025] Extract the deviation between the predicted value of prior synchronization error and the actual physical observation timestamp;

[0026] The deviation correction amount of the actual physical observation timestamp deviation relative to the prior synchronization error prediction value is calculated using the Kalman filter gain matrix variables.

[0027] The deviation correction is accumulated to the prior synchronization error prediction value to obtain the optimal posterior state feedback value, which is then formatted and encapsulated as the synchronization error scalar signal.

[0028] The specific processing logic for extracting bidirectional long short-term memory network features from historical bandwidth sequence data to obtain the predicted bandwidth boundary vector includes:

[0029] The historical bandwidth sequence data is input into the cell state matrix along the forward and reverse channels respectively. The forward output tensor and the reverse output tensor are extracted and concatenated to generate a full-dimensional hidden feature matrix.

[0030] Apply geometric projection mapping matrix weights to the scalar values ​​of the output feature vectors at each time step of the full-dimensional hidden feature matrix, and perform exponential distribution normalization calculation to obtain the attention weights for the corresponding time steps;

[0031] The predicted bandwidth boundary vector is obtained by weighting and summing the scalar values ​​of the output feature vector according to the attention weights.

[0032] As a preferred technical solution of the present invention, the nonlinear penalty mapping scaling process of the synchronization error scalar signal to obtain the dynamic synchronization constraint factor includes the following specific processing logic: multiplying the synchronization error scalar signal by the physical error amplification scaling factor, and performing an exponential mapping operation with the natural constant as the base to generate the dynamic synchronization constraint factor.

[0033] Furthermore, the joint prediction bandwidth boundary vector and dynamic synchronization constraint factor perform reward function correction interaction calculations on the evaluation network embedded in the dual-delay deep deterministic policy gradient network. The specific processing logic includes:

[0034] Obtain initial values ​​for basic network action value assessment;

[0035] Extract the values ​​of specific dimension elements corresponding to downlink bandwidth quota allocation from the resource allocation action vector, calculate the squared difference between the value and the predicted bandwidth boundary vector, subtract the fault tolerance bandwidth dynamic redundancy constant safety threshold, and multiply by the bandwidth over-limit absolute penalty weight constant to obtain the bandwidth penalty term.

[0036] Multiply the dynamic synchronization constraint factor by the synchronization failure fatal penalty weight constant to obtain the synchronization failure penalty term;

[0037] Subtracting the bandwidth penalty and the synchronization failure penalty from the initial value of the basic network action value assessment yields the joint overall reward quantification value.

[0038] The model strategy is updated by backpropagation using the negative gradient of the quantified value of the joint total reward.

[0039] As a preferred embodiment of the present invention, after outputting the resource allocation action vector for controlling the physical edge nodes, the specific execution steps of the termination determination logic include:

[0040] Continuously calculate the norm geometric distance difference between the transient resource allocation action vector of adjacent hardware interrupt cycles and the historical resource allocation action vector;

[0041] When the geometric distance difference of the norm is determined to be less than the factory-configured action convergence limit small decision constant, and the transient optimal posterior state feedback value is lower than the system rigid tolerance limit, a forced interruption control signal is sent downstream and the memory of the relevant neural network hidden feature temporary storage area is cleared.

[0042] Furthermore, the step of inputting the joint matrix of system environment state into a dual-delay deep deterministic policy gradient network for global policy inference, implemented based on a multi-agent federated distillation architecture, specifically includes:

[0043] Deploy a local policy branch distribution mapping network structure within the physical edge node on the base station side to perform pre-feature inference, and only output the dimensionality reduction probability parameter feature value tensor to the cloud central aggregation node;

[0044] The cloud-based central aggregation node calculates the probability distribution divergence tensor between the global policy backbone distribution mapping network structure and the local policy branch distribution mapping network structure.

[0045] The communication quality absolute trust weight constant derived from the physical communication connection error rate is used to calculate the physical value of the global distillation loss, and the backpropagation gradient flow generated by the physical value of the global distillation loss is used to update the tensor parameters of the backbone global network matrix.

[0046] As a preferred embodiment of the present invention, the process of adjusting the hardware computing power slice boundary of the computing node according to the resource allocation action vector and dynamically refreshing the hardware encoding quantization parameters of the video stream embeds an adaptive forward error correction redundancy dynamic allocation mechanism based on the underlying physical environment. The specific processing logic includes:

[0047] Extract the synchronization error scalar signal from the end-to-end timestamp alignment signal, and parse the available bandwidth margin scalar between the predicted bandwidth boundary vector and the current physical network card interface transient true throughput scalar;

[0048] Construct a forward error correction redundancy dynamic scaling equation, perform joint cross-domain feature mapping on the available bandwidth margin scalar and the synchronization error scalar signal, and calculate the target error correction block generation ratio for a single physical data link.

[0049] Based on the target error correction block generation ratio, redundant check data packets are directly inserted into the network communication data link layer protocol stack hardware pipeline of the physical edge node;

[0050] The specific mathematical expression of the forward error correction redundancy dynamic scaling equation is as follows:

[0051] ;

[0052] In the formula, The target error correction block generation ratio is calculated and output, and is dimensionless. A scalar of basic forward error correction redundancy coefficients preset for the underlying hardware driver; The synchronization error scalar signal is in milliseconds. Sensitivity mapping multiplier for the redundancy state of the underlying network physical bandwidth; The available bandwidth margin is a scalar value, in megabits per second; It is a numerical scalar value for evaluating the priority status of a specific concurrent terminal in the service flow priority status vector.

[0053] As a preferred embodiment of the present invention, the specific execution steps of adjusting the hardware computing power slicing boundary of the computing node, which integrates a physical topology-level affinity resource locking mechanism based on a non-uniform memory access architecture, specifically include:

[0054] The hardware topology jump constant between the physical media access control layer hardware interface of the peripheral component interconnect bus on the physical edge node motherboard and the independent L3 cache space of each specific central processing unit core is detected.

[0055] Based on the extracted service flow priority state vector and the hardware topology hop constant, a memory access latency cost tensor matrix spanning non-uniform memory access architecture node regions is constructed.

[0056] Based on the gradient descent trend of the elements inside the memory access latency cost tensor matrix, the target processor core index feature parameter with a global minimum value is found. The kernel scheduling instruction is used to forcibly migrate the computation location of the media data packet forwarding core thread with high service flow priority and bind it to the hardware processing pipeline corresponding to the target processor core index feature parameter, triggering the direct cache access action of bypass memory.

[0057] The mathematical calculation model for the local elements within the memory access latency cost tensor matrix is ​​as follows:

[0058] ;

[0059] In the formula, In order to target the The concurrent business data stream is allocated to the first... The scalar value of memory access latency cost generated per CPU core; The physical delay baseline multiplier for hardware transmission; For physical network card slot modules up to the first The hardware topology jump constant for each central processing unit core is a positive integer; To calculate the priority penalty for resource transfer, convert the weight scalar. For the service flow priority state vector, the one for the first A numerical scalar representing the priority status assessment of a concurrent business data stream; A very small positive floating-point constant set to prevent division by zero exceptions.

[0060] In addition, the present invention also discloses a cloud conferencing audio and video resource dynamic allocation optimization system enabled by edge computing, comprising a first signal acquisition and evaluation extraction module, a second signal acquisition and prediction compensation module, a matrix fusion and penalty mapping module, and a joint policy inference and hardware control module; the system is configured to execute the cloud conferencing audio and video resource dynamic allocation optimization method enabled by edge computing as described above.

[0061] Compared with the prior art, the present invention has the following beneficial effects:

[0062] This invention achieves deterministic allocation of computing resources in complex edge network environments by constructing a physical closed-loop feedback system that integrates cross-level, multi-dimensional features. First, it performs time-decrease attention priority evaluation processing on multi-dimensional perceptual interaction signals, quantifying the real-time attention weights of different terminals. This accurately separates silent audio streams or inactive video streams, allowing physical computing power to be directed towards data streams with high interaction value and reducing ineffective power consumption. Simultaneously, it performs range standardization processing on node fluctuation signals, uniformly mapping the physical dimensions of heterogeneous hardware indicators to a standardized numerical space. This eliminates the feature overwhelming effect of huge values ​​on weak signals in tensor operations, enhancing the accuracy of the system's assessment of physical node states.

[0063] This invention performs Kalman filtering clock drift compensation on end-to-end timestamp alignment signals, filtering out Gaussian white noise interference in the physical transmission channel and outputting a denoised synchronization error scalar signal, providing a stable benchmark for audio and video time alignment. It also applies bidirectional long short-term memory network feature extraction to historical bandwidth sequence data to generate a high-confidence predicted bandwidth boundary vector, transforming reactive bandwidth allocation into forward-looking prediction and reducing the probability of transmission congestion caused by bandwidth abrupt changes.

[0064] Building upon this foundation, the service flow priority state vector and the multi-dimensional resource state vector are concatenated into a joint system environment state matrix. Furthermore, the synchronization error scalar signal is nonlinearly mapped into a dynamic synchronization constraint factor, which is then injected into a dual-delay deep deterministic policy gradient network. By combining the predicted bandwidth boundary vector and the dynamic synchronization constraint factor to perform corrective interactive calculations on the reward function of the evaluation network, the reinforcement learning network is forced to strictly adhere to the inviolable physical bandwidth and clock synchronization limits when generating resource allocation action vectors that adjust hardware computing power slice boundaries and video stream hardware encoding quantization parameters. This avoids the risk of system availability dropping to zero due to the policy network blindly pursuing throughput extrema, thereby eliminating the meaningless trial-and-error losses of physical computing power while ensuring the determinism and stability of cloud conferencing transmission.

[0065] This invention introduces a physical topology-level affinity resource locking mechanism based on NUMA architecture, which forcibly binds the processing threads of high-priority data streams to the optimal processor core and bypasses the system's main memory, significantly improving cache hit rate and reducing cross-node bus load and access latency. By introducing an adaptive forward error correction redundancy dynamic allocation mechanism, the error correction redundancy ratio is dynamically adjusted according to synchronization error and bandwidth margin, realizing differentiated anti-packet loss protection for different priority data streams using scarce bandwidth resources, further improving the audio and video synchronization accuracy and system survivability of cloud conferencing under harsh network conditions. Attached Figure Description

[0066] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0067] Figure 1 This is a flowchart illustrating the overall process of the method of the present invention.

[0068] Figure 2 This is a flowchart of the termination determination logic of the present invention.

[0069] Figure 3 This is a flowchart of the adaptive forward error correction redundancy dynamic allocation mechanism of the present invention. Detailed Implementation

[0070] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the embodiments of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0071] The following is in conjunction with the appendix Figures 1-3 The embodiments of the present invention will be described in detail below.

[0072] Example 1: This example discloses a cloud conferencing audio and video resource dynamic allocation optimization system enabled by edge computing, which includes a first signal acquisition and evaluation extraction module, a second signal acquisition and prediction compensation module, a matrix fusion and penalty mapping module, and a joint policy inference and hardware control module.

[0073] This embodiment also details a method for dynamically allocating and optimizing cloud conferencing audio and video resources powered by edge computing. In practice, the execution entity is a dedicated acceleration control chip embedded within the user plane functional network element of the physical edge computing node in 5G mobile communication technology. The entire physical execution cycle is strictly controlled by a microsecond-level hardware clock interrupt generator.

[0074] Specifically as follows:

[0075] This embodiment quantifies the terminal interaction intent of the application layer and the physical hardware resource status of the underlying layer in real time and maps them uniformly to a standardized numerical space.

[0076] Step 1-1: Acquisition of multidimensional sensory interaction signals and assessment of time-decayed attention;

[0077] In cloud conferencing scenarios, physical edge nodes often fail to identify the interaction intentions of key physical terminals, leading to an even distribution of physical computing power across different data streams and resulting in ineffective computing power usage. The specific technical problem to be solved in step 1-1 is: how to quantify the real-time attention weights of different concurrent terminals, thereby separating out silent data streams with no interactive value.

[0078] The multidimensional sensory interaction signals specifically include a hand-raising microphone snatching signaling packet initiated by the end-user hardware and a screen-sharing request packet establishing a transmission control protocol connection. The first signal acquisition and evaluation extraction module calls the floating-point arithmetic unit inside the dedicated acceleration control chip to apply time-attenuation attention priority evaluation processing to the multidimensional sensory interaction signals. The first signal acquisition and evaluation extraction module incorporates basic mathematical calculation rules; the calculation logic for the evaluation processing is presented by the following formula:

[0079] ;

[0080] In the formula, Representing the The priority status evaluation value of each concurrent user, without dimensionless units; Represents the basic permission weight coefficient of a network role, with no dimensionless unit; This represents the attention decay rate control coefficient, in units of... ; Representing the The physical time interval between the current processor state and the last time the multi-dimensional sensing interaction signal was triggered for each concurrent user is a variable, with units of [missing information]. ; A constant representing the total number of physical interaction types; a positive integer. Representative and the The system gain correction coefficient corresponding to each type of physical interaction, with no unit of measurement; This is a logical Boolean indicator variable representing whether an action has been triggered, with a value of 0 or 1.

[0081] For example, the first concurrent user is designated as a regular participant, and a basic network role permission weight coefficient is set. Set the attention time decay rate control coefficient to 1.2. It is 0.5 ; Retrieve the physical time interval since the last triggered interaction action. It is 2.0. Set the total number of physical interaction types as a constant. The value is 2, where the first type is hand-raising signaling and the second type is screen-sharing signaling; set the corresponding system gain correction coefficient. It is 3.0. Version 5.0; The system detected that the first concurrent user had just triggered a hand-raising action but had not yet triggered screen sharing, therefore a logical Boolean indicator variable was extracted. =1, The result is 0. Substituting the above value into the exponential decay term, we obtain... Calculate the gain accumulation term to obtain... Adding the exponential decay term to the gain accumulation term yields the priority state evaluation value. The first signal acquisition and evaluation extraction module performs calculations on all concurrently accessing physical terminals in the current conference room and outputs a service flow priority state vector in one-dimensional floating-point array format.

[0082] Attention time decay rate control coefficient and system gain correction factor These parameters are not arbitrarily set. They were determined by collecting 500 hours of standard cloud meeting log data, extracting the speaker switching frequency and the physical interval between terminal signaling triggers, and then performing polynomial data fitting using the least squares method on an offline training server. The parameter selection aligns with the statistical law of temporal decay in human attention shifts during video conferences. This computational logic solves the technical problem of lacking differentiated labeling for multimodal concurrent data streams, enabling physical edge nodes to dynamically tilt their computing power based on the physical interaction intent of the terminal.

[0083] Step 1-2: Multidimensional feature extraction and standardization of node fluctuation signals;

[0084] After identifying the application layer business weights, the system faces the constraint of the constantly fluctuating physical state of the physical edge nodes themselves. The physical dimensions and numerical ranges of heterogeneous computing resources (such as CPU frequency, memory paging rate, and network card queues) are completely different, and direct input will cause serious deviations in the gradient updates of subsequent neural networks.

[0085] The first signal acquisition and evaluation extraction module extracts node fluctuation signals by reading parameters from the server motherboard's built-in system management bus registers. These parameters include the CPU core's physical thread instruction cycle utilization, memory random access page swapping frequency, and network interface card (NIC) physical queue backlog depth. This module employs range standardization logic to perform multi-dimensional feature extraction. The specific range standardization mathematical model is as follows:

[0086] ;

[0087] In the formula, Representing the The dimensionless mapping values ​​obtained after standardization of the physical characteristics of the item nodes; Representing the The actual transient sampling value of the physical characteristics of the item node at the current moment, with the unit determined according to the specific physical characteristics; Representing the The lower limit reference constant of the physical characteristics of the item node; Representing the The upper limit threshold constant of the physical characteristics of the item node.

[0088] For example, setting the current real transient sample value of the physical queue backlog depth of the network interface card. For 850 data packets; extract the lower limit baseline constant for the physical queue backlog depth of the network interface card. For 0 packets; extract the upper limit threshold constant for the physical queue backlog depth of the network interface card. There are 1024 data packets. Substituting the values ​​into the numerator yields... Substituting the numerical values ​​into the denominator yields... Performing a division operation yields a dimensionless mapping value. The first signal acquisition and evaluation extraction module traverses all acquisition dimensions and generates a multi-dimensional resource state vector whose values ​​are strictly mapped to the interval between 0 and 1.

[0089] The lower limit reference constant and upper limit threshold constant of the node physical characteristics are absolute physical extreme values ​​determined by the hardware factory calibration file and the physical edge node full-load stress test report, ensuring that the range normalization process has a rigid boundary at the physical level. This range normalization processing logic solves the technical problem of inconsistent dimensions of monitoring indicators in heterogeneous hardware and eliminates the feature smothering effect of extremely large values ​​on weak signals in tensor operations.

[0090] Step 2-1: Kalman clock drift compensation for end-to-end timestamp alignment signals;

[0091] After being routed and forwarded through multiple layers of network nodes, audio and video packets will experience time-series stacking errors when they reach the core router. The physical transmission link is also subject to Gaussian white noise interference; directly extracting timestamps from packets will cause severe physical jitter at the receiving end, making it unsuitable as a time alignment reference.

[0092] The second signal acquisition and prediction compensation module intercepts the end-to-end timestamp alignment signal from the physical receiver port located between the data link layer and the network layer in the network protocol stack. The second signal acquisition and prediction compensation module then initiates Kalman filter clock drift compensation processing. The operation initializes the system state transition matrix in the memory pool. The mathematical model for its Kalman filter gain update and posterior state estimation is as follows:

[0093] ;

[0094] ;

[0095] In the formula, Represents time series moments The Kalman filter gain matrix variable is dimensionless. Represents time series moments The prior physical error covariance matrix; The mathematical matrix representing the observation mapping; The adjoint transpose of the mathematical matrix representing the observation mapping; Represents the measurement noise covariance matrix; This represents the optimal posterior state feedback value extracted after removing Gaussian noise, i.e., the synchronization error scalar signal, in units of... ; This represents the predicted value of the prior synchronization error, in units of... ; Represents time series moments Extracted real physical observation timestamp deviation, in units of .

[0096] For example, setting the prior physical error covariance value in a one-dimensional state. Set to 5.0; define the mathematical scalar for observation mapping. The value is 1.0, and its transpose matrix scalar Both are set to 1.0; set the measurement noise covariance value. Set to 2.0; set the prior synchronization error prediction value. It is 15.0 ; Set the actual physical observation timestamp deviation of the network interface It is 18.0 Perform Kalman filter gain calculation; the part within parentheses is calculated as follows. .

[0097] After calculating and inverting, we get Multiplying this by the numerator yields the Kalman filter gain matrix variable. The optimal posterior state feedback value is updated, and the deviation residual is calculated as follows: The correction amount is calculated based on the gain. The corrections are accumulated to the prior predictions to obtain the synchronization error scalar signal. .

[0098] Measurement noise covariance matrix in Kalman filtering algorithm The variance of the basic white noise was determined by long-term data collection under no-load conditions at the physical receiving port, reflecting the inherent interference baseline of the physical transmission channel. The aforementioned Kalman filter clock drift compensation logic solves the problem of severe timestamp jitter caused by network congestion, outputting a smooth and physically accurate synchronization error scalar signal, providing a core reference for the precise alignment of subsequent time-dimension resources.

[0099] Step 2-2: Bidirectional Long Short-Term Memory Network Prediction Based on Historical Bandwidth Sequence Data

[0100] In dynamic cloud conferencing environments, the physical throughput of communication at the underlying network routing interface exhibits dramatic nonlinear tidal fluctuations. Traditional reactive bandwidth allocation strategies always lag behind changes in the physical channel state, leading to severe packet loss and crashes in audio and video streams when encountering sudden congestion.

[0101] The second signal acquisition and prediction compensation module continuously monitors the network interface controller register to obtain the physical throughput of the underlying network routing interface communication, assembling it into a continuous time-series signal. This continuous time-series signal is input into the bidirectional long short-term memory (LSTM) network feature extraction processing flow. Within the cell state matrix of the bidirectional LSM network, the continuous time-series signal undergoes forward and reverse channel input operations, respectively. The second signal acquisition and prediction compensation module performs an absolute concatenation operation on the forward and reverse output tensors in the feature channel dimension to generate a full-dimensional hidden feature matrix. This full-dimensional hidden feature matrix is ​​then pushed into the weight layer of a time-attention mechanism network configured with an exponentially distributed normalized activation function to perform geometric projection dimensionality reduction, calculating the prediction bandwidth boundary vector. The mathematical logic of this geometric projection is as follows:

[0102] ;

[0103] In the formula, This represents the prediction bandwidth boundary vector after fusing temporal attention weights, in units of... ; The constant representing the total number of time steps for extracting historical throughput sequences is a positive integer; The weights of the geometric projection mapping matrix representing the fully connected layer; Represents the full-dimensional hidden feature matrix at the th The output feature vector scalar value at each time step, in units of ; Represents the full-dimensional hidden feature matrix at the th The output feature vector scalar value at each time step, in units of .

[0104] In practical implementation, for example, a constant is set for the total number of timing steps. The step size is 2. Extract the scalar values ​​of the feature vector output from the network layer. It is 80.0 , It is 100.0 Extract the scalar corresponding to the weights of the geometric projection mapping matrix of the fully connected layer. The result is 0.02. An exponential mapping operation is performed on the first time step, yielding... Perform an exponential mapping operation on the second time step to obtain... Calculate the sum of the normalized exponents in the denominator: Calculate the attention weights for the first time step: Calculate the attention weights for the second time step: The final solution calculates the prediction bandwidth boundary vector. .

[0105] The geometric projection mapping matrix weights of the fully connected layer used in the above formula The parameters are fixed by extracting actual tidal fluctuation log data of a 10 Gigabit fiber optic backbone network over three months from an offline data center and training iteratively through 100,000 backpropagation iterations using an adaptive moment estimation optimizer.

[0106] Bidirectional temporal feature extraction combined with exponential normalization calculation solves the technical problem that unidirectional prediction cannot detect bandwidth mutation trends, and outputs a downlink physical throughput boundary reference with high confidence.

[0107] Step 3: System joint matrix fusion and dynamic synchronization factor mapping;

[0108] Multimodal decision-making requires complete alignment of environmental features within the mathematical space. The matrix fusion and penalty mapping module receives data in the cache channel and executes direct concatenation logic of memory address blocks. This module performs feature matrix concatenation and alignment processing on the service flow priority state vector and the multidimensional resource state vector. It also uses zero-padding operations to fill in missing tensor dimensions, generating a joint system-environment state matrix representing the system's physical overall state. Simultaneously, the module extracts the synchronization error scalar signal and applies a nonlinear penalty mapping scaling process containing natural logarithm base mapping features. This nonlinear penalty mapping scaling process exponentially amplifies the minute jitter of the error scale within the mathematical space, generating a dynamic synchronization constraint factor that correlates with the critical point of physical quality collapse. Specifically, the mathematical expression of this nonlinear penalty mapping scaling process is:

[0109] ;

[0110] In the formula, This refers to the generated dynamic synchronization constraint factor; This is the scaling factor for physical error amplification; The synchronization error scalar signal is expressed in milliseconds. This factor exponentially amplifies minute synchronization errors, enabling subsequent reinforcement learning networks to detect the critical risk of synchronization degradation.

[0111] Step 4: Inference and reward correction using a dual-delay deep deterministic policy gradient network;

[0112] Existing reinforcement learning-based scheduling models often blindly pursue the maximum overall throughput, ignoring the risk of zero availability due to physical synchronization failures. Step 4 aims to address the technical problem of how to force the neural network to adhere to inviolable physical bandwidth and clock synchronization limits when exploring high-throughput actions.

[0113] The joint policy inference and hardware control module initiates the deployment of a tensor processing core component with a dual-delay deep deterministic policy gradient network architecture. This module uses the joint system environment state matrix as the tensor input layer node data for global policy inference processing. It jointly predicts the bandwidth boundary vector and uses the dynamic synchronization constraint factor to perform reward function correction and interactive calculations on the two sets of mirror twin structure evaluation networks embedded within the dual-delay deep deterministic policy gradient network. The core physical constraint fusion logic is as follows:

[0114] ;

[0115] In the formula, This represents the total quantitative value of the joint reward; The representative evaluation network is used for the joint matrix of system environment state. and resource allocation action vectors The initial value of the basic network action value is obtained through predictive calculations; The absolute penalty weight constant representing bandwidth exceeding the limit; Represents the resource allocation action vector The numerical values ​​of specific dimension elements corresponding to the downlink bandwidth quota allocation are in units of ; Represents the predicted bandwidth boundary vector, in units of ; This represents the fault-tolerant bandwidth dynamic redundancy constant safety threshold. The constant representing the fatal penalty weight for synchronization failure; This represents the scaling factor for physical error amplification; Represents a scalar signal of synchronization error, in units of .

[0116] The reward function calculation was validated based on the data generated in the previous steps. Initial values ​​for the basic network action value assessment were extracted. Set to 50.0; extract the absolute penalty weight constant for bandwidth out-of-bounds access. The value is 0.5; extract sub-dimension truncation values. It is 120.0 Substitute the previously calculated prediction bandwidth boundary vector. Extract the fault-tolerant bandwidth dynamic redundancy constant safety threshold. Set to 100.0; Extract the fatal penalty weight constant for synchronization failure. 2.0; Extract the physical error amplification scaling factor. The value is 0.1; substitute it into the scalar signal of the synchronization error calculated in the previous steps. .

[0117] Bandwidth penalty calculation:

[0118] The square of the difference is .because The value is Multiply by the penalty weight to get Synchronization failure penalty calculation: Calculate the exponential part. Multiply by the fatal penalty weight to get Finally, the combined total reward is quantified. If a severely negative reward occurs, the backpropagation process will generate a huge amount of gradient updates, forcing the actuator network to abandon the current overbandwidth allocation of action vectors.

[0119] Weight constants in the model , , as well as It is a physical extreme value limit constant set by the slope of the derivative of the packet loss rate curve recorded by the basic performance test bench when it is subjected to an injection-based network blocking attack.

[0120] The underlying control daemon parses the finally converged resource allocation action vector structure and directly sends a set of control instructions to the operating system kernel. This set of instructions adjusts the hardware computing power slice boundaries of the computing nodes bound to each processing thread and modifies the constant bitrate parameter configuration of the video encoding library. The reward constraint calculation logic solves the technical problem of the policy network lacking physical boundary awareness, eliminating the meaningless trial-and-error losses of physical computing power while ensuring high availability.

[0121] In parallel, the system initiates a termination decision logic based on norm geometric distance: the joint strategy inference and hardware control module continuously calculates the Euclidean norm 2 distance difference between the transient resource allocation action vector of adjacent hardware interrupt cycles and the historical resource allocation action vector; when the distance difference is less than the factory-configured action convergence limit small decision constant, and the latest synchronization error scalar signal exceeds the system's rigid tolerance limit, a forced interrupt control signal is triggered, the memory of the neural network hidden feature temporary storage area is cleared, and the computing system is forced to enter a sleep waiting period to prevent deadlock.

[0122] Example 2: This example is a further optimization based on Example 1. In this example, the data closed-loop interaction logic and deadlock prevention mechanism between modules are further explained.

[0123] The system prevents data congestion deadlocks caused by external network outages or excessively high hardware temperatures triggering computing power lock-up. A dual-threshold termination logic based on a Euclidean norm 2 geometric distance calculation model is deployed within the physical pipeline. This logic continuously calculates the norm geometric distance difference between the transient resource allocation action vectors of adjacent hardware interrupt cycles and the historical resource allocation action vectors.

[0124] The transient resource allocation action vector is extracted as an array [120.0, 50.0] containing two dimensions, and the historical resource allocation action vector is extracted as [120.1, 49.9].

[0125] The norm geometric distance difference is calculated as follows:

[0126] .

[0127] The built-in convergence limit of the factory action extraction system is a small constant of 0.2, and the rigid tolerance limit of the extraction system is 25.0. The decision logic compares 0.141 with 0.2, confirms that it is less than a constant, and extracts the optimal posterior state feedback value as 17.142. The result was confirmed to be below the rigid tolerance limit. No full-module computation interruption response was triggered. This occurred when the motion vector converged and the time error amplified dramatically, exceeding 25.0. At this point, the physical judgment logic triggers a forced control signal to be sent downstream, clearing the memory of the neural network's hidden feature temporary storage area, forcing the computing system to enter a dormant waiting period to await temperature or channel recovery. The dual-threshold termination judgment logic eliminates the computing power dead loop blocking under extreme physical scenarios.

[0128] Example 3: This example is based on the same principles as Example 1, except that it discloses a multi-agent distributed evolution architecture for collaborative computing deployment of massive edge micro base stations.

[0129] The network topology of the joint policy inference and hardware control module comprises a single cloud-based central aggregation node and at least one base station-side physical edge node. The architecture executes distributed evolutionary execution logic based on multi-agent federated distillation. Each physical edge node deploys an independent local policy branch distribution mapping network structure. After performing policy inference locally, this local network extracts and outputs only a dimensionality-reduced probability parameter feature value tensor representing the action probability distribution. The cloud-based central aggregation node receives the dimensionality-reduced probability parameter feature value tensors uploaded by each physical edge node and calculates the Kourbach-Leibler divergence tensor between the global probability distribution output by its global policy backbone distribution mapping network structure and each local probability distribution.

[0130] The algorithm introduces an absolute trust weight constant for communication quality based on the bit error rate of the physical communication connection at each physical edge node. The divergence tensors of each local node are summed with the trust weights, and the global distillation loss is calculated as a physical value through a weighted summation. An exemplary calculation method is as follows:

[0131] ;

[0132] In the formula, The physical loss of global distillation is expressed in numerical form. This represents the total number of physical edge nodes. For the first The absolute trust weight constant for the communication quality of each physical edge node. Calculate the function for the Kourbak-Leibler divergence. The probability distribution output by the network structure is mapped to the global policy backbone distribution. For the first The local policy branch distribution of each physical edge node maps to the probability distribution of the network structure output. The cloud-based central aggregation node uses the backpropagation gradient flow generated by the global distillation loss physics to synchronously update the tensor parameters of the backbone global network matrix. The updated quantized mapping weights are then redistributed to each physical edge node.

[0133] The distributed evolutionary architecture completely cuts off the network load on the cloud physical communication fiber core caused by the direct convergence of high-dimensional matrices, maintaining the high availability and physical fault tolerance of the system's cross-regional node computing resource scheduling.

[0134] To facilitate a better understanding of the present invention by those skilled in the art, the present invention will be further illustrated below with reference to specific implementation examples.

[0135] In practical implementation, the system is set to be deployed in an edge computing network element of a fifth-generation mobile communication backbone network, carrying an ultra-high-definition cross-domain cloud conferencing process containing 150 concurrent nodes. During this transient microsecond-level physical clock interruption period, the first signal acquisition and evaluation extraction module captures the signaling data packet of the 15th terminal user triggering the raising of hands to seize the microphone through bypass port mirroring.

[0136] The system extracts the user's network role basic permission weight coefficient, setting it to 1.5, and obtains the system-set attention time decay rate control coefficient as 0.4. The physical timer reports that the physical time interval variable between the user's current processor time and the last time the multi-dimensional perception interaction signal was triggered is 1.2 seconds. Since only the hand-raising signal was triggered, the system extracts the system gain correction coefficient associated with the hand-raising action as 4.0, and the logical Boolean indicator variable as 1; the screen sharing action was not triggered, and the corresponding logical Boolean indicator variable was extracted as 0. The first signal acquisition and evaluation extraction module substitutes the above data into the time decay attention calculation model, first calculating the exponential decay term: Multiply by the network role's basic permission weighting coefficient. ,get The gain accumulation term is calculated as follows: Finally, the two values ​​are added together to obtain the priority status assessment value for the 15th terminal concurrent access user. The first signal acquisition and evaluation extraction module traverses all 150 concurrent terminals and generates a service flow priority state vector consisting of 150 floating-point numbers.

[0137] Within a synchronous microsecond timeframe, the first signal acquisition and evaluation extraction module calls the system management bus register to extract the physical characteristics of the physical edge node. Currently, the main chip of the physical edge node is under high load, and the actual transient sampled value of the CPU core physical thread instruction cycle utilization is 88. By consulting the chip's manufacturer's technical specifications, the lower limit benchmark constant for extracting this physical characteristic is 5, and the upper limit threshold constant is 100.

[0138] The first signal acquisition and evaluation extraction module uses range standardization processing logic to calculate... Obtain the molecule and calculate. Obtain the denominator. Perform division to obtain the dimensionless mapping value. Range standardization processes parameters such as memory page swapping rate and network interface card (NIC) physical queue depth, combining them to generate a multi-dimensional resource state vector. Through range standardization, the fluctuation range of various heterogeneous hardware data is strictly compressed and mapped to a purely mathematical range of 0 and 1, completely avoiding the feature overwhelming of tiny values ​​such as NIC queue depth by huge values ​​such as cache frequency in subsequent tensor fusion.

[0139] Upon entering the underlying network synchronization observation phase, the second signal acquisition and prediction compensation module intercepts the end-to-end timestamp alignment signal in the network protocol stack. The system initializes the prior physical error covariance of the single-dimensional Kalman filter to 4.5, the observation mapping mathematical scalar to 1.0, and the measurement noise covariance to 2.5 based on real-time physical channel background interference assessment. Based on the recursive state of the previous cycle, the system extracts a prior synchronization error prediction value of 18.5 milliseconds, and simultaneously obtains the actual physical observation timestamp deviation carried by the current data packet as 24.5 milliseconds through protocol stack parsing. The Kalman filter gain matrix variables are calculated: the denominator is calculated. ;molecular ; divide to get Then, the optimal posterior state feedback value is updated, and the deviation between the actual observation value and the prior prediction value is calculated: Milliseconds; Calculate the deviation correction amount: Milliseconds. The deviation correction is accumulated to the prior prediction value of 18.5 milliseconds, and the system outputs a synchronization error scalar signal. Milliseconds. This filtering process forcibly filters out the random extreme value jitter caused by Gaussian white noise in the physical channel.

[0140] In the bandwidth trend prediction stage, the second signal acquisition and prediction compensation module injects the extracted hidden features from two consecutive time steps into the cell state matrix of the bidirectional long short-term memory network. The scalar value of the output feature vector at the first time step is set to 45.0 megabits per second, and at the second time step to 120.0 megabits per second. The weights of the geometric projection mapping matrix of the fully connected layer of the temporal attention network are extracted to 0.015. Exponential distribution normalization calculation is performed.

[0141] First step length exponent term Calculate the natural index ;

[0142] The second step length exponent term Calculate the natural index .

[0143] Sum of normalized denominators: The attention weights at time step 1 are calculated as follows: Weights at the second time step: The weights and feature scalars are cross-weighted and summed. , The final output prediction bandwidth boundary vector is Megabits per second.

[0144] In the cross-dimensional data fusion and inference stage, the matrix fusion and penalty mapping module performs hard concatenation of priority and resource state tensors within the cache. The dual-latency deep deterministic policy gradient network outputs a transient resource allocation action vector, which includes a downlink bandwidth quota control instruction truncation value set at 145.0 megabits per second. The system performs rigorous physical red line review calculations on this policy. The initial value of the basic network action evaluation is extracted to be 150.0. Bandwidth penalty calculation is then performed.

[0145] Calculate the difference between the allocation intention 145.0 and the prediction boundary 101.505: Squared difference: ; Safety threshold after deducting fault-tolerant bandwidth dynamic redundancy constant: Multiply by the absolute penalty weight constant for bandwidth out-of-bounds access: This leads to the bandwidth penalty term.

[0146] Perform synchronization penalty calculation: multiply the synchronization error scalar signal by the error amplification scaling factor. ; Calculate the natural index: Multiply by the synchronization failure fatal penalty weight constant: This leads to the synchronization failure penalty term. Finally, the reward constraint logic intervenes: calculating the quantified value of the joint overall reward. Faced with deeply negative reward feedback, the reinforcement learning twin evaluation network immediately triggers backpropagation in the negative gradient direction, forcibly discarding the super-resolution action that could lead to potential physical congestion. In subsequent iterations, it converges and outputs a safe resource allocation action vector with downlink quota compressed to 95.0 megabits per second. The underlying control daemon parses this vector and decentralizes it to the hardware layer to adjust the concurrent threads of the general-purpose graphics processor and compress the data stream encoding of black-screen terminals. The system is entirely based on deterministic computation within mathematical space, preventing the risk of network-wide collapse caused by blindly overclocking for throughput, and achieving precise data stream allocation within physical constraints.

[0147] Example 4: This example is a further optimization based on Example 1. In this example, an independent thermal state sensing mechanism is added to solve the technical problem of frequency reduction caused by thermal accumulation.

[0148] In long-term high-concurrency audio and video encoding / decoding scenarios, thermal accumulation in the chip transistors of physical edge nodes can cause a sharp drop in the physical clock frequency, resulting in the instantaneous failure of resource state features in the system environment state joint matrix. This embodiment adds an independent thermal state sensing mechanism to solve this technical problem.

[0149] The matrix fusion and penalty mapping module extracts the thermal gradient parameters of the chip transistors at physical edge nodes in real time. The computation process incorporates the following thermal blocking equation:

[0150] ;

[0151] In the formula, The generated thermal blocking penalty factor; The preset absolute penalty weight constant for thermal overshooting; This is the thermodynamic sudden change sensitivity coefficient; This is the transient physical temperature value of the chip transistor read from the motherboard thermistor, in degrees Celsius. This is the safety temperature baseline constant, expressed in degrees Celsius. This is the extreme temperature constant at which the logic of a silicon crystal malfunctions, expressed in degrees Celsius.

[0152] For example: Extracting the safety temperature baseline constant The collapse temperature extreme constant is 65.0 degrees Celsius. The temperature is 95.0 degrees Celsius. This is the current transient physical temperature of the chip's transistors. The temperature soared to 85.0 degrees Celsius. The sensitivity coefficient to thermal sudden change was extracted. The absolute penalty weight constant for thermal out-of-bounds thermal exceedance is 3.0. The value is 50.0. Calculate the normalized thermodynamic approximation rate: Perform exponentiation: Solve for the thermal blocking penalty factor. .

[0153] In the joint policy inference stage, the dual-delay deep deterministic policy gradient network greatly reduces the thermal blocking penalty factor. The resource allocation action vector is directly deducted from the overall reward, forcing the network to actively reduce the frame rate or abandon high-resolution encoding. This algorithm mechanism interacts with the bandwidth prediction mechanism. When the network bandwidth is sufficient but thermal collapse is imminent, the blocking algorithm forcibly suppresses the throughput limit, ensuring the physical survivability of physical edge nodes under thermal limits.

[0154] Example 5: This example is based on the same principles as Example 1, except that it addresses the issue of cross-modal timestamp separation and loss of synchronization.

[0155] The human visual and auditory centers have extremely low tolerance for lip-sync errors. Synchronization constraints fail to distinguish between audio and video streams in the physical coding pipeline. This embodiment addresses the issue of cross-modal timestamp separation synchronization loss.

[0156] The second signal acquisition and prediction compensation module performs independent Kalman filter clock drift compensation processing on the audio and video timing at the receiving port, and outputs audio synchronization error scalar signals respectively. Scalar signal with video synchronization error The matrix fusion and penalty mapping module performs cross-residual nonlinear penalty mapping scaling processing, and its interactive correlation expression is as follows:

[0157] ;

[0158] In the formula, This refers to the cross-modal dynamic synchronization constraint factor. This is the extreme value weight constant for cross-modal failure; It is the hyperbolic tangent activation function; For residual excitation multipliers; This is a scalar signal representing audio synchronization error, measured in milliseconds. This is a scalar signal representing video synchronization error, measured in milliseconds. It is the critical constant for cross-modal physical tolerance that is perceptible to human physiology, expressed in milliseconds.

[0159] For example, extracting the audio synchronization error scalar signal. The time interval is 12.0 milliseconds, representing the scalar signal of video synchronization error caused by the loss of reference frames. The time reached 82.0 milliseconds. The absolute difference was 70.0 milliseconds. The critical constant for cross-modal physical tolerance was extracted. The residual excitation multiplier is 40.0 milliseconds. The extreme value weight constant for cross-modal failure is 0.1. The value is 100.0. Calculate the residual base: Milliseconds. Hyperbolic tangent mapping calculation: Solving for the cross-modal dynamic synchronization constraint factor. This factor is injected into the reinforcement learning evaluation network as the highest priority penalty term. Under this constraint, the policy network generates resource allocation action vectors that will enforce extreme video structure similarity target degradation instructions. By issuing quantization parameters that sacrifice image quality to the video encoder, it forces the video processing pipeline to quickly catch up with the audio timestamp.

[0160] Physical environment: 150 concurrent nodes, 15% bursty background packet loss, and forced setting of base station cooling fans to stop for 3 minutes.

[0161] Table 1: Comparison of the effects of extended algorithm technology;

[0162] Average node response latency (milliseconds) 125.4 42.1 38.6 Peak temperature of the edge chip core (degrees Celsius) 98 (Triggered a system crash) 89 78 (Stable Inhibition) Probability of labial slip-synchronization (greater than 40 milliseconds) 14.8% 5.2% 0.3% Ineffective computing power consumption ratio 31% 6% 4.5%

[0163] As shown in Table 1, after supplementing the thermodynamic and cross-modal interaction equations, the system has achieved a leapfrog objective physical progress in survival stability and audio-visual synchronization accuracy under extreme conditions.

[0164] Example 6: Based on the infrastructure described in Example 1, this example further discloses a specific implementation of an adaptive forward error correction redundancy dynamic allocation mechanism for differentiated packet loss protection on unreliable physical channels.

[0165] Physical signal fading and routing congestion in wireless air interfaces or cross-border fiber optic links can lead to the loss of random data packets. Conventional retransmission mechanisms require initiating retransmission requests at the application layer, and their round-trip delays can disrupt the timestamp alignment of real-time audio and video. Therefore, the system needs to pre-allocate redundancy at the data link layer to combat channel packet loss.

[0166] The joint policy inference and hardware control module extracts the predicted bandwidth boundary vector generated by the bidirectional long short-term memory network within a microsecond-level scheduling cycle. The difference between the current physical network interface's transient actual throughput scalar and the actual throughput scalar is calculated as the available bandwidth margin scalar. Simultaneously, acquire the synchronization error scalar signal generated in the preceding steps. and the priority status evaluation numerical scalar of the current concurrent terminals. .

[0167] The system calls the hardware-level floating-point arithmetic unit to load the forward error correction redundancy dynamic scaling equation:

[0168] ;

[0169] In the formula, The target error correction block generation ratio represents the solution output and is dimensionless. The scalar represents the preset basic forward error correction redundancy coefficient. This represents a scalar signal indicating synchronization error, measured in milliseconds. Sensitivity mapping multipliers representing the redundancy state of the underlying network physical bandwidth; This represents a scalar value indicating available bandwidth margin, measured in megabits per second. This represents a numerical scalar value indicating the priority status assessment of the corresponding terminal.

[0170] In practical implementation, for example, setting a prediction bandwidth boundary vector Mbps, the current throughput is 80.0 Mbps, Mbps. Priority for highly active user #15. , ms, get , Calculate the denominator: , denominator Its reciprocal is Calculate the logarithmic term: .final This means that for this terminal, approximately 19.1% of redundant error correction data needs to be added in addition to the basic payload.

[0171] For silent terminal number 42, Recalculate the logarithmic terms With other parameters remaining unchanged, we get The redundancy ratio was reduced to 1.02%.

[0172] The underlying network control daemon, based on the target error correction block generation ratio of each terminal, issues instructions to the field-programmable gate array (FPGA) network interface card (NIC) accelerator. This instructs the NIC to perform Reed-Solomon coding on the media data blocks at the data link layer, generate check packets according to the calculated ratio, and insert them into the transmission queue. The receiving end hardware decoder then uses these check packets for packet loss recovery, avoiding application layer retransmissions. This mechanism prioritizes scarce bandwidth and computing resources for high-value data streams, suppressing the unnecessary channel consumption caused by low-priority data streams.

[0173] In some preferred embodiments, a resource locking mechanism based on the physical topology hierarchy affinity of the Non-Unified Memory Access (NUMA) architecture is further disclosed, which is integrated into the hardware control process of step 4.

[0174] Multi-slot physical edge servers suffer from cross-node memory access latency penalties. The operating system may allocate network card interrupt requests to remote central processing unit cores, triggering a large amount of cross-chip interconnect bus data movement, causing bus contention and latency jitter.

[0175] The first signal acquisition and evaluation module calls the system NUMA application programming interface to probe the physical topology. It obtains the physical topology jump constants from the current physical network card slot to each central processing unit core. For example, the core hop count in node zero, which is directly connected to the network card, is 0, while the core hop count in node one, which is connected to other nodes, is 1.

[0176] The joint policy inference and hardware control module constructs a memory access latency cost tensor matrix for each concurrent data stream. and candidate core Calculate the cost scalar:

[0177] ;

[0178] In the formula, Representing the The data stream is allocated to the first... A core cost scalar; The physical delay baseline multiplier for hardware transmission; This is the physical topology jump constant; To calculate the priority penalty for resource transfer, convert the weight scalar. For the first The priority evaluation value for each data stream; To prevent extremely small positive floating-point constants from being divided by zero.

[0179] For example: setting , , User #15 The cost of assigning to the core of node zero on the same side (skip 0): The cost of allocating to the core of remote node 1 (jump 1): The cost gradient is significant, and the system locks the ipsilateral core as the optimal target by finding the global minimum.

[0180] For silent terminal number 42 The cost of allocating to the near-end core is The remote core is Both are extremely large, indicating that low-priority services are not sensitive to physical location.

[0181] Based on this, the system issues thread affinity instructions to the operating system kernel for high-priority data streams, forcibly binding the corresponding media forwarding threads to the optimal core selected in the cost matrix, and modifying the network card's direct memory access descriptor to write data directly to the corresponding L3 cache area of ​​that core, bypassing the system's main memory. This simplifies the physical layer of the data path, reducing cross-node bus load and access latency.

[0182] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0183] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. It should be noted that any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for dynamically allocating and optimizing cloud conferencing audio and video resources powered by edge computing, characterized in that, Includes the following steps: Step 1: Collect multi-dimensional perception interaction signals from the cloud conferencing platform, perform time-decay attention priority evaluation processing on the multi-dimensional perception interaction signals to obtain a business flow priority state vector; simultaneously collect node fluctuation signals from the edge computing physical environment, perform multi-dimensional feature extraction processing on the node fluctuation signals to obtain a multi-dimensional resource state vector. Step 2: Acquire the end-to-end timestamp alignment signal of the physical receiving port, perform Kalman filtering clock drift compensation processing on the end-to-end timestamp alignment signal to obtain the synchronization error scalar signal; synchronously acquire the historical bandwidth sequence data of the underlying network link, perform bidirectional long short-term memory network feature extraction processing on the historical bandwidth sequence data to obtain the predicted bandwidth boundary vector; Step 3: Perform feature matrix concatenation and alignment processing on the service flow priority state vector and the multidimensional resource state vector to obtain the system environment state joint matrix; perform nonlinear penalty mapping scaling processing on the synchronization error scalar signal to obtain the dynamic synchronization constraint factor. Step 4: Input the joint matrix of system environment state into the dual-delay deep deterministic policy gradient network for global policy inference processing. Combine the predicted bandwidth boundary vector and the dynamic synchronization constraint factor to perform reward function correction interaction calculation on the evaluation network embedded in the dual-delay deep deterministic policy gradient network. Output resource allocation action vector for controlling physical edge nodes. Adjust the hardware computing power slice boundary of computing nodes according to the resource allocation action vector and dynamically refresh the hardware encoding quantization parameters of video stream, and trigger the termination judgment logic based on norm geometric distance.

2. The edge computing-enabled cloud conferencing audio and video resource dynamic allocation optimization method according to claim 1, characterized in that, The process of performing time-decay attention priority evaluation on the multi-dimensional perceptual interaction signal to obtain the business flow priority state vector includes the following specific processing logic: Extract the physical time interval variable and logical Boolean indicator variable for concurrent users; The physical time interval variable is substituted into the preset attention time decay rate control coefficient to perform an exponential decay operation, resulting in an exponential decay term; The logical Boolean indicator variable is combined with the corresponding system gain correction coefficient and an accumulation operation is performed to obtain the gain accumulation term; By integrating the network role basic permission weight coefficient, the exponential decay term, and the gain accumulation term, a priority status evaluation value is obtained; and a one-dimensional floating-point array format of the service flow priority status vector is generated by traversing all concurrent access physical terminals.

3. The edge computing-enabled cloud conferencing audio and video resource dynamic allocation optimization method according to claim 1, characterized in that, The process of extracting multidimensional features from node fluctuation signals to obtain a multidimensional resource state vector includes the following specific processing logic: Extract the true transient sample values, lower limit reference constant, and upper limit threshold constant of various node physical characteristics from the node fluctuation signal; The difference between the actual transient sample value and the lower limit reference constant is calculated as the numerator; The difference between the upper limit threshold constant and the lower limit benchmark constant is calculated as the denominator; Divide the numerator by the denominator to obtain a dimensionless mapping value; The multidimensional resource state vector is generated by combining the dimensionless mapping values ​​of all collected dimensions.

4. The edge computing-enabled cloud conferencing audio and video resource dynamic allocation optimization method according to claim 1, characterized in that, The end-to-end timestamp alignment signal is subjected to Kalman filtering clock drift compensation processing to obtain a synchronization error scalar signal. The specific processing logic includes: Extract the prior physical error covariance matrix, the observation mapping mathematical matrix, and the measurement noise covariance matrix, and calculate the Kalman filter gain matrix variables; Extract the deviation between the predicted value of prior synchronization error and the actual physical observation timestamp; The deviation correction amount of the actual physical observation timestamp deviation relative to the prior synchronization error prediction value is calculated using the Kalman filter gain matrix variables. The deviation correction is accumulated to the prior synchronization error prediction value to obtain the optimal posterior state feedback value, which is then formatted and encapsulated as the synchronization error scalar signal.

5. The edge computing-enabled cloud conferencing audio and video resource dynamic allocation optimization method according to claim 1, characterized in that, The process of extracting bidirectional long short-term memory network features from historical bandwidth sequence data to obtain the predicted bandwidth boundary vector includes the following specific processing logic: The historical bandwidth sequence data is input into the cell state matrix along the forward and reverse channels respectively. The forward output tensor and the reverse output tensor are extracted and concatenated to generate a full-dimensional hidden feature matrix. Apply geometric projection mapping matrix weights to the scalar values ​​of the output feature vectors at each time step of the full-dimensional hidden feature matrix, and perform exponential distribution normalization calculation to obtain the attention weights for the corresponding time steps; The predicted bandwidth boundary vector is obtained by weighting and summing the scalar values ​​of the output feature vector according to the attention weights.

6. The edge computing-enabled cloud conferencing audio and video resource dynamic allocation optimization method according to claim 1, characterized in that, The process of performing nonlinear penalty mapping scaling on the synchronization error scalar signal to obtain the dynamic synchronization constraint factor includes the following specific processing logic: multiplying the synchronization error scalar signal by the physical error amplification scaling factor, and performing an exponential mapping operation with the natural constant as the base to generate the dynamic synchronization constraint factor.

7. The edge computing-enabled cloud conferencing audio and video resource dynamic allocation optimization method according to claim 1, characterized in that, The joint prediction bandwidth boundary vector and dynamic synchronization constraint factor perform reward function correction and interactive calculation on the evaluation network embedded in the dual-delay deep deterministic policy gradient network. The specific processing logic includes: Obtain initial values ​​for basic network action value assessment; Extract the values ​​of specific dimension elements corresponding to downlink bandwidth quota allocation from the resource allocation action vector, calculate the squared difference between the initial value of basic network action value assessment and the predicted bandwidth boundary vector, subtract the fault tolerance bandwidth dynamic redundancy constant security threshold, and multiply by the absolute penalty weight constant for bandwidth overrun to obtain the bandwidth penalty term. Multiply the dynamic synchronization constraint factor by the synchronization failure fatal penalty weight constant to obtain the synchronization failure penalty term; Subtracting the bandwidth penalty and the synchronization failure penalty from the initial value of the basic network action value assessment yields the joint overall reward quantification value. The model strategy is updated by backpropagation using the negative gradient of the quantified value of the joint total reward.

8. The edge computing-enabled cloud conferencing audio and video resource dynamic allocation optimization method according to claim 1, characterized in that, After the output is used to control the resource allocation action vector of the physical edge node, the specific execution steps of the termination determination logic include: Continuously calculate the norm geometric distance difference between the transient resource allocation action vector of adjacent hardware interrupt cycles and the historical resource allocation action vector; When the geometric distance difference of the norm is determined to be less than the factory-configured action convergence limit small decision constant, and the transient optimal posterior state feedback value is lower than the system rigid tolerance limit, a forced interruption control signal is sent downstream and the memory of the relevant neural network hidden feature temporary storage area is cleared.

9. The edge computing-enabled cloud conferencing audio and video resource dynamic allocation optimization method according to claim 1, characterized in that, The process of inputting the joint matrix of system environment states into a dual-delay deep deterministic policy gradient network for global policy inference, implemented based on a multi-agent federated distillation architecture, specifically includes: Deploy a local policy branch distribution mapping network structure within the physical edge node on the base station side to perform pre-feature inference, and only output the dimensionality reduction probability parameter feature value tensor to the cloud central aggregation node; The cloud-based central aggregation node calculates the probability distribution divergence tensor between the global policy backbone distribution mapping network structure and the local policy branch distribution mapping network structure. The communication quality absolute trust weight constant derived from the physical communication connection error rate is used to calculate the physical value of the global distillation loss, and the backpropagation gradient flow generated by the physical value of the global distillation loss is used to update the tensor parameters of the backbone global network matrix.

10. The edge computing-enabled cloud conferencing audio and video resource dynamic allocation optimization method according to claim 1, characterized in that, The process of adjusting the hardware computing power slice boundary of the computing node according to the resource allocation action vector and dynamically refreshing the hardware encoding quantization parameters of the video stream embeds an adaptive forward error correction redundancy dynamic allocation mechanism based on the underlying physical environment. The specific processing logic includes: Extract the synchronization error scalar signal from the end-to-end timestamp alignment signal, and parse the available bandwidth margin scalar between the predicted bandwidth boundary vector and the current physical network card interface transient true throughput scalar; Construct a forward error correction redundancy dynamic scaling equation, perform joint cross-domain feature mapping on the available bandwidth margin scalar and the synchronization error scalar signal, and calculate the target error correction block generation ratio for a single physical data link. Based on the target error correction block generation ratio, redundant check data packets are directly inserted into the network communication data link layer protocol stack hardware pipeline of the physical edge node.

Images