A network communication quality real-time prediction system and method based on digital twinning

Abstract

The application discloses a network communication quality real-time prediction system and method based on digital twinning, relates to the technical field of network communication, and comprises a physical network end, a data management end, a twinned network end, a cooperative evaluation end, a strategy execution end and a synchronous monitoring module; the physical network end collects network operation state data through active detection and passive sniffing and converts the network operation state data into a standardized state vector; the twinned network end constructs a digital twinning model based on real-time and historical data, outputs a future communication quality prediction value through multi-time scale simulation deduction, and compares and calculates a quality correlation coefficient by comparing an actual state with the prediction value; the acquisition frequency and simulation parameters are dynamically adjusted according to the quality correlation coefficient, and a closed-loop feedback is formed. Through the bidirectional cooperation of the physical network and the twinned network and the parameter self-adaptive mechanism, the application solves the technical problem that the prediction deviation of the existing scheme is accumulated and cannot be corrected in a closed loop.

Description

Technical Field

[0001] This invention relates to the field of network communication technology, and in particular to a real-time prediction system and method for network communication quality based on digital twins. Background Technology

[0002] With the rapid development of 5G, IoT, and cloud computing technologies, communication networks are continuously expanding in scale and becoming increasingly complex, carrying a wider variety of services. Various emerging applications are placing increasingly stringent demands on network bandwidth, latency, capacity, security, and reliability. Against this backdrop, network operation and maintenance models are shifting from traditional reactive maintenance to proactive predictive maintenance, making real-time prediction of network communication quality crucial for ensuring business continuity and user experience.

[0003] Digital twin technology, due to its ability to perform large-scale, high-fidelity simulations in virtual environments, is gradually becoming an effective means of solving real-world network challenges. In the field of communication networks, digital twin technology is used to construct virtual models of networks to simulate and predict network behavior, performance, and changes.

[0004] However, existing technical solutions generally lack an effective two-way collaboration and closed-loop feedback mechanism between the physical network and the twin network. Specifically, existing solutions typically only collect data from the physical network and input it into the twin model for simulation prediction, but the prediction results are difficult to feed back to the physical network in real time to guide its operational status adjustment. At the same time, the simulation parameters of the twin model are usually fixed once set, and cannot be dynamically and adaptively adjusted according to the deviation between the actual state of the physical network and the prediction results. When the communication quality of the physical network fluctuates drastically, the twin model cannot promptly detect the prediction deviation and adjust its own parameters, causing the prediction results to gradually deviate from the actual network state, and the prediction accuracy to continuously decline over time. This unidirectional, open-loop architecture makes it difficult for the system to continuously provide highly reliable communication quality prediction results in complex and ever-changing network environments.

[0005] Therefore, there is an urgent need for a real-time prediction system and method for network communication quality that can achieve bidirectional collaboration between physical networks and twin networks and has closed-loop adaptive capabilities. Summary of the Invention

[0006] The main objective of this invention is to provide a real-time prediction system and method for network communication quality based on digital twins, so as to solve the problems mentioned in the background above.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: a real-time prediction system and method for network communication quality based on digital twins, comprising a physical network terminal, a data governance terminal, a twin network terminal, a collaborative evaluation terminal, and a policy execution terminal, wherein the physical network terminal, the data governance terminal, the twin network terminal, the collaborative evaluation terminal, and the policy execution terminal are all equipped with a synchronous monitoring module.

[0008] The physical network terminal is used to collect the operating status data of each network node in the physical communication network in real time, and convert the collected raw data into a standardized state vector through data preprocessing; the physical network terminal is also used to receive quality prediction feedback from the collaborative evaluation terminal, and adjust the data collection frequency and collection granularity according to the feedback information.

[0009] The twin network terminal is used to construct a digital twin model of the physical communication network based on the real-time status data and historical operation data of the physical network terminal. It then inputs the current physical network state vector into the digital twin model for simulation and outputs predicted network communication quality values ​​for multiple future time windows.

[0010] The collaborative evaluation terminal is used to compare the state vector actually collected by the physical network terminal with the state vector predicted by the twin network terminal after each prediction period, calculate the quality correlation coefficient used to characterize the prediction accuracy, and issue collection strategy adjustment instructions to the physical network terminal and simulation inference parameter adjustment instructions to the twin network terminal based on the quality correlation coefficient, thus forming a closed-loop collaborative mechanism from data collection at the physical network terminal to simulation inference at the twin network terminal and then to deviation detection and feedback adjustment at the collaborative evaluation terminal.

[0011] The synchronous monitoring module is used to monitor in real time the data acquisition delay at the physical network end, the simulation and deduction time at the twin network end, and the changing trend of the quality correlation coefficient at the collaborative evaluation end. When the data acquisition delay exceeds a preset threshold, the acquisition strategy is adjusted. When the simulation and deduction time exceeds a preset threshold, the model simplification mechanism is triggered. When the quality correlation coefficient is continuously lower than a preset threshold, an early warning signal is issued.

[0012] Furthermore, the operational status data of the physical network includes link bandwidth utilization, end-to-end latency, packet loss rate, jitter value, and signal strength.

[0013] The physical network terminal includes a data acquisition module, a data preprocessing module, and a feedback execution module. The data acquisition module includes an active detection unit and a passive sniffing unit. The active detection unit sends probe packets to target nodes in the network, measuring end-to-end latency and packet loss rate through the round-trip time and packet loss of the probe packets. The passive sniffing unit acquires real-time traffic data on network links through port mirroring or traffic mirroring, and statistically analyzes link bandwidth utilization and traffic distribution characteristics. The data preprocessing module includes a data cleaning unit and a feature extraction unit. The data cleaning unit performs outlier detection and feature extraction on the acquired raw data. The data collection process includes data removal, missing value imputation, and time alignment operations. The feature extraction unit extracts network communication quality feature vectors from the cleaned data. These feature vectors include bandwidth utilization sequences, latency sequences, packet loss rate sequences, and jitter value sequences for each link, and these sequences are concatenated to form a standardized state vector. The feedback execution module receives quality prediction feedback from the collaborative evaluation terminal and dynamically adjusts the data collection frequency based on the prediction confidence index in the feedback: when the prediction confidence is lower than a preset threshold, the data collection frequency and probe packet sending density are increased; when the prediction confidence is higher than the preset threshold, the data collection frequency is reduced to decrease network overhead.

[0014] Furthermore, the data governance terminal is used to uniformly store, manage versions, and verify the quality of multi-source heterogeneous operating status data collected by the physical network terminal, construct a time-series dataset containing historical status data and real-time status data, and provide a standardized data access interface for the twin network terminal; the data governance terminal is also used to dynamically adjust the retention strategy and sampling compression rate of historical data according to the monitoring results of the synchronization monitoring module.

[0015] The data governance module includes a storage management module, a version control module, and a data service module. The storage management module includes a time-series database unit and a compression and archiving unit. The time-series database unit stores standardized state vectors collected from the physical network in a time-series manner, with each data record containing a timestamp, node identifier, and a complete feature vector. The compression and archiving unit dynamically adjusts the sampling compression rate of historical data based on the monitoring results of the synchronization monitoring module, increasing the compression rate when the storage I / O load exceeds a preset threshold and decreasing the compression rate when the data query response time exceeds a preset threshold. The version control module manages the training dataset of the digital twin model, creating a new data version snapshot each time the twin network initiates a model update, and retaining a preset number of recent versions for rollback and comparison. The data service module includes a standardized interface unit and a permission management unit. The standardized interface unit provides a unified data query API to the twin network, supporting data retrieval by time range, node range, and feature dimension. The permission management unit controls access permissions for different data users to the dataset.

[0016] Furthermore, the twin network terminal includes a topology twin module, a behavioral twin module, and a projection and prediction module; the topology twin module includes a topology discovery unit and a topology update unit; the topology discovery unit is used to identify the connection relationships between nodes in the physical network through a network layer topology discovery protocol, generating a network topology graph containing a set of nodes and a set of edges; the topology update unit is used to detect topology changes based on real-time status data from the physical network terminal, and triggers incremental updates of the topology graph when changes in link connectivity or routing are detected; the behavioral twin module includes a traffic modeling unit and a channel modeling unit; the traffic modeling unit is used to establish a traffic arrival model based on historical traffic data provided by the data governance terminal. The traffic arrival model employs a self-similar traffic model or a Markov-modulated Poisson process model to generate traffic arrival patterns in the simulation. The channel modeling unit establishes a channel quality model, which calculates link-level channel attenuation and bit error rate based on signal strength, interference level, and signal-to-noise ratio parameters. The prediction module takes the standardized state vector at the current moment as initial conditions and inputs it into the digital twin model. It then performs simulation and deduction using discrete event simulation or network calculus-based analysis methods, outputting a predicted state vector sequence for the next T time windows. This predicted state vector sequence includes the predicted bandwidth utilization, predicted end-to-end delay, predicted packet loss rate, and predicted jitter value for each link.

[0017] Furthermore, the inference and prediction module adopts a multi-timescale prediction strategy: on a short timescale, a deterministic analysis method based on network calculus is used to predict the communication quality within 1 to 5 time windows in the future, with each time window being 100ms long; on a long timescale, a time-series prediction model trained on historical data is used to predict the communication quality within 6 to 20 time windows in the future, with each time window being 1s long; the short-timescale prediction results and the long-timescale prediction results are weighted and fused to generate the final prediction value, and the fusion weight is dynamically adjusted according to the stability of the current network state, increasing the weight of the short-timescale prediction when the network state fluctuates significantly.

[0018] Furthermore, the correlation calculation module is used to, at the end of each prediction period, convert the data actually collected by the physical network terminal... Time-state vector and Siamese network endpoints Time prediction By comparing the state vectors at different times, the quality correlation coefficient is calculated using the following formula:

[0019] in, for The mass correlation coefficient at time t, its value range is: ; is the dimension of the state vector; The first data actually collected at the physical network end Dimensional state value; The k-th dimension state value predicted by the twin network endpoint; For the first The weighting coefficients of the dimensional state variables are preset according to the importance of each dimension to communication quality; This is the similarity function.

[0020] When the quality correlation coefficient When the value is greater than the preset threshold of 0.85, the current prediction of the Siamese network is considered to have high accuracy; when... When the value is below the preset threshold of 0.60, it is determined that the prediction deviation of the twin network is too large, triggering the parameter adjustment mechanism.

[0021] The parameter adaptive module is used to dynamically adjust the simulation and inference parameters of the Siamese network based on the quality correlation coefficient, specifically including: when When the value is below 0.60, shorten the prediction time window length. And increase the precision parameters of the simulation; when When the value is higher than 0.90, the prediction time window length T is increased and the simulation refinement parameter is reduced to decrease computational overhead; when Keep the current parameter unchanged when the value is between 0.60 and 0.90.

[0022] The confidence generation module is used to generate a predicted confidence index based on the time series of the quality correlation coefficient, as shown in the following formula:

[0023] in, for The prediction confidence index at time t, with a value range of . ; The quality correlation coefficient at the current moment; For the most recent The moving average of the mass correlation coefficient at each time point; For the most recent Normalized standard deviation of the quality correlation coefficient at each time point; For the weighting coefficients, satisfying .

[0024] Furthermore, the strategy execution terminal is used to receive the network communication quality prediction value and prediction confidence index output by the collaborative evaluation terminal, convert the prediction result into network scheduling instructions or traffic optimization actions according to the preset strategy rules, and feed the execution result back to the synchronous monitoring module to verify the strategy effect.

[0025] The policy execution end includes a policy parsing module, an instruction conversion module, and an effect verification module. The policy parsing module receives the network communication quality prediction value and prediction confidence index output by the collaborative evaluation end, and matches the corresponding execution policy according to a preset policy rule base, which contains several mapping rules. The instruction conversion module converts the matched execution policy into standard configuration instructions or control signaling for specific network devices. The standard configuration instructions include routing adjustment instructions, bandwidth limiting instructions, and queue scheduling instructions. The effect verification module, after the instruction conversion module issues the execution action, obtains the network status change data after execution through the synchronous monitoring module, compares the actual change effect with the predicted change effect, generates a policy execution effect report, and feeds it back to the collaborative evaluation end.

[0026] Furthermore, the synchronization monitoring module includes a latency monitoring unit, a resource scheduling unit, and an anomaly warning unit. The latency monitoring unit is used to collect data acquisition latency at the physical network end, storage I / O latency at the data governance end, simulation and deduction time at the twin network end, and instruction issuance latency at the policy execution end every preset period. The resource scheduling unit is used to trigger a model simplification mechanism when the simulation and deduction time exceeds a preset threshold. The model simplification mechanism includes reducing the simulation spatial resolution, reducing the simulation time granularity, or enabling a fast simulation mode. When the storage I / O latency exceeds a preset threshold, it triggers the adjustment of the compression and archiving strategy at the data governance end. The anomaly warning unit is used to send a warning signal to the network management system when the quality correlation coefficient is lower than the preset threshold of 0.50 for three consecutive periods, indicating that there is a significant deviation between the digital twin model and the physical network, and suggesting that the model be recalibrated.

[0027] Furthermore, the active detection unit sends detection messages using the UDP-based TWAMP protocol or ICMP timestamp request messages; the active detection unit dynamically adjusts the sending frequency of detection messages and the number of target nodes based on the prediction confidence fed back by the collaborative evaluation end. When the prediction confidence is lower than 0.60, the detection frequency is increased to twice the original frequency, and the number of target nodes is increased by 50%.

[0028] A method for real-time prediction of network communication quality based on digital twins includes the following steps: S1: The data acquisition module at the physical network end collects the operating status data of each network node in the physical communication network through active detection and passive sniffing. The data preprocessing module converts the collected raw data into a standardized state vector.

[0029] S2: The data governance end performs unified storage and version management of multi-source heterogeneous operation status data collected from the physical network end, and constructs a time-series dataset containing historical status data and real-time status data, providing a standardized data access interface for the twin network end.

[0030] S3: The topology twin module constructs and continuously updates the network topology map based on the real-time status data provided by the data governance terminal, while the behavior twin module establishes a traffic arrival model and a channel quality model based on historical traffic data and channel data.

[0031] S4: The simulation and prediction module takes the standardized state vector at the current moment as the initial condition input to the digital twin model, performs simulation and deduction through discrete event simulation or network calculation analysis methods, and outputs the predicted values ​​of network communication quality for multiple future time windows.

[0032] S5: At the end of each prediction cycle, the correlation calculation module of the collaborative evaluation end compares the actual state vector collected by the physical network end with the predicted state vector by the twin network end, and calculates the quality correlation coefficient. .

[0033] S6: The parameter adaptive module is based on the quality correlation coefficient. The prediction time window length and simulation granularity parameters of the twin network are dynamically adjusted. The confidence generation module generates a prediction confidence index based on the time series of the quality correlation coefficient and sends an acquisition strategy adjustment command to the physical network.

[0034] S7: The policy execution end receives the network communication quality prediction value and prediction confidence index output by the collaborative evaluation end, and converts the prediction results into network scheduling instructions or traffic optimization actions according to the preset policy rules and issues them for execution.

[0035] S8: The synchronous monitoring module monitors the running latency and performance indicators of each terminal in real time. When the latency exceeds the preset threshold, the corresponding adjustment mechanism is triggered. When the quality correlation coefficient is continuously lower than the preset threshold, an early warning signal is issued.

[0036] The present invention has the following beneficial effects: 1. In this invention, by setting up a physical network terminal and deploying a data acquisition module that combines active detection and passive sniffing within it, as well as a feedback execution module that dynamically adjusts the acquisition frequency according to the prediction confidence, the invention achieves the function of uniformly converting multi-source heterogeneous raw data into a standardized state vector and making the data acquisition granularity adaptively change with the prediction deviation. This solves the technical problem in existing solutions where the fixed data acquisition frequency leads to sparse data or excessive network overhead during critical periods.

[0037] 2. In this invention, by setting up a data governance terminal and configuring time-series database storage, version management, and a unified data service interface, it plays a role in orderly storage and version control of multi-source state data and provides a standardized access channel for twin models. This solves the technical problems of scattered historical and real-time data and chaotic model training dataset versions in existing solutions, which make rollback difficult.

[0038] 3. In this invention, by setting up a twin network terminal and configuring topological twin, behavioral twin, and multi-timescale inference and prediction modules, it plays a role in constructing a dynamically updatable digital network image in virtual space and balancing short-term accuracy with long-term trends in simulation and inference. This solves the technical problems in existing solutions where twin models cannot reflect dynamic changes in topology and single-timescale prediction is difficult to meet actual needs.

[0039] 4. This invention establishes a collaborative evaluation terminal and deploys correlation calculation, parameter adaptation, and confidence generation modules within it. The correlation calculation module calculates the quality correlation coefficient between the actual state and the predicted state at the end of each prediction cycle. The parameter adaptation module simultaneously issues adjustment instructions for the acquisition frequency and simulation parameters to both the physical network terminal and the twin network terminal based on this coefficient. This effectively establishes a closed-loop collaborative mechanism from data acquisition to simulation deduction to deviation detection and feedback correction, solving the technical problem in existing solutions where the lack of bidirectional collaboration between the physical network and the twin network leads to the accumulation of prediction deviations over time and the inability to adaptively correct them.

[0040] 5. This invention sets up a policy execution terminal and configures policy parsing, instruction conversion, and effect verification modules. The policy parsing module matches preset policy rules based on predicted values ​​and confidence levels. The instruction conversion module converts the policy into standard configuration instructions for network devices and issues them for execution. The effect verification module compares the actual changes with the predicted changes and generates an execution report, which is fed back to the collaborative evaluation terminal. This invention automatically converts the prediction results into executable network control actions and performs closed-loop verification of their effects, solving the technical problems in existing solutions where prediction results cannot automatically drive network configuration adjustments and policy effects lack tracking and verification.

[0041] 6. This invention sets up a synchronous monitoring module and configures delay monitoring, resource scheduling, and anomaly early warning units. The delay monitoring unit periodically collects the running delay of each terminal. The resource scheduling unit triggers model simplification or compression strategy adjustment when simulation time or storage I / O exceeds the threshold. The anomaly early warning unit issues model calibration warning when the quality correlation coefficient is continuously lower than the threshold. This plays a role in uniformly monitoring the operating status of each terminal of the system and triggering active intervention when performance deteriorates. It solves the technical problems of existing solutions that lack a global monitoring perspective and cannot deal with prediction failures or resource overloads in a timely manner. Attached Figure Description

[0042] Figure 1 This is a schematic diagram of the overall system architecture of a real-time network communication quality prediction system and method based on digital twins according to the present invention; Figure 2 This is a schematic diagram of the physical network architecture of the network communication quality real-time prediction system and method based on digital twins according to the present invention. Figure 3 This is a schematic diagram of the data governance terminal of a real-time prediction system and method for network communication quality based on digital twins according to the present invention. Figure 4 This is a schematic diagram of the architecture of the twin network end of the real-time prediction system and method for network communication quality based on digital twins according to the present invention. Figure 5 This is a schematic diagram of the collaborative evaluation terminal of a real-time prediction system and method for network communication quality based on digital twins according to the present invention. Figure 6 This is a schematic diagram of the strategy execution terminal of the real-time prediction system and method for network communication quality based on digital twins according to the present invention. Figure 7 This is a schematic diagram of the architecture of the synchronous monitoring module of a real-time prediction system and method for network communication quality based on digital twins according to the present invention. Detailed Implementation

[0043] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0044] Please refer to Figure 1As shown, this invention provides a real-time network communication quality prediction system based on digital twins, comprising a physical network terminal, a data governance terminal, a twin network terminal, a collaborative evaluation terminal, and a policy execution terminal. These terminals share a common synchronous monitoring module. The physical network terminal is deployed at the edge node of the physical communication network or on the central network management server. It establishes data connections with each network node in the physical communication network through a network interface to collect real-time operational status data of each network node. This operational status data includes link bandwidth utilization, end-to-end latency, packet loss rate, jitter value, and signal strength. In terms of intensity, the physical network end preprocesses the collected raw data into a standardized state vector, and then forwards this state vector to the twin network end via the data governance end. The physical network end also receives quality prediction feedback from the collaborative evaluation end through the data governance end, and adjusts the data collection frequency and granularity based on the feedback information. The data governance end establishes bidirectional data channels with both the physical network end and the twin network end for unified storage, version management, and quality verification of the multi-source heterogeneous operational status data collected by the physical network end, constructs a time-series dataset containing historical and real-time status data, and provides a standardized data access interface for the twin network end. The twin network endpoint establishes a data reading connection with the data governance endpoint to construct a digital twin model of the physical communication network based on real-time status data from the physical network endpoint and historical operational data provided by the data governance endpoint. It then simulates and extrapolates the current physical network state vector into the digital twin model, outputting predicted network communication quality values ​​for multiple future time windows. These predicted values ​​are forwarded to the policy execution endpoint via the collaborative evaluation endpoint. The collaborative evaluation endpoint establishes bidirectional data connections with the physical network endpoint, data governance endpoint, twin network endpoint, and policy execution endpoint, respectively, to compare the actual state vector collected by the physical network endpoint with the predicted state vector from the twin network endpoint at the end of each prediction cycle. The parameters are compared, and a quality correlation coefficient is calculated to characterize the accuracy of the prediction. Based on this quality correlation coefficient, a data acquisition strategy adjustment command is issued to the physical network end and a simulation inference parameter adjustment command is issued to the twin network end, forming a closed-loop collaborative mechanism from data acquisition at the physical network end to simulation inference at the twin network end and then to deviation detection and feedback adjustment at the collaborative evaluation end. Data connections are established at the strategy execution end, the collaborative evaluation end, and the synchronous monitoring module to receive the network communication quality prediction value and prediction confidence index output by the collaborative evaluation end. According to the preset strategy rules, the prediction results are converted into network scheduling commands or traffic optimization actions, and the execution results are fed back to the synchronous monitoring module.The synchronous monitoring module establishes monitoring data connections with the physical network, data governance, twin network, collaborative evaluation, and policy execution ends, respectively. This is used to monitor in real time the data acquisition latency of the physical network, the storage I / O latency of the data governance end, the simulation and deduction time of the twin network, the trend of the quality correlation coefficient at the collaborative evaluation end, and the instruction issuance latency at the policy execution end. When the data acquisition latency exceeds a preset threshold, the acquisition strategy is adjusted; when the simulation and deduction time exceeds a preset threshold, a model simplification mechanism is triggered; and when the quality correlation coefficient continuously falls below a preset threshold, an early warning signal is issued.

[0045] Please refer to Figure 2As shown, the physical network end includes a data acquisition module, a data preprocessing module, and a feedback execution module. The data acquisition module includes an active detection unit and a passive sniffing unit. The output of the active detection unit establishes a data connection with the input of the data cleaning unit of the data preprocessing module. The active detection unit sends TWAMP protocol or ICMP timestamp request messages based on UDP to target nodes in the network. By measuring the round-trip time and packet loss of the detection messages, it obtains end-to-end latency data and packet loss rate data. The active detection unit also dynamically adjusts the sending frequency of detection messages and the number of target nodes to be detected based on the prediction confidence fed back from the collaborative evaluation end. When the prediction confidence is lower than 0.60, the detection frequency is increased to... The frequency is doubled and the number of target nodes detected is increased by 50%. The output of the passive sniffing unit establishes a data connection with the input of the data cleaning unit of the data preprocessing module. The passive sniffing unit obtains real-time traffic data on the network link through port mirroring or traffic mirroring, and statistically analyzes the bandwidth utilization and traffic distribution characteristics of each link from this real-time traffic data. The data preprocessing module includes a data cleaning unit and a feature extraction unit. The input of the data cleaning unit establishes a data connection with the output of the active detection unit and the output of the passive sniffing unit, respectively. The data cleaning unit performs outlier detection and removal, missing value imputation, and time alignment operations on the collected raw data. Outlier detection adopts a 3x frequency... The Laida criterion method for standard deviation is used, and missing value imputation employs linear or spline interpolation. Time alignment synchronizes data from different sources according to a unified timestamp. A data connection is established between the output of the data cleaning unit and the input of the feature extraction unit. The feature extraction unit extracts network communication quality feature vectors from the cleaned data. These feature vectors include bandwidth utilization sequences, latency sequences, packet loss rate sequences, and jitter value sequences for each link. The feature extraction unit concatenates these sequences according to a preset dimensional order to form a standardized state vector, which is a floating-point array of dimension K. The output of the data preprocessing module is connected to the twin network via a push mechanism. The prediction module establishes a data connection and transmits the standardized state vector to the twin network. The input of the feedback execution module establishes a data connection with the confidence generation module of the collaborative evaluation end through the data governance end. The feedback execution module receives the quality prediction feedback issued by the collaborative evaluation end. This quality prediction feedback includes a prediction confidence index. The feedback execution module dynamically adjusts the data acquisition frequency according to the prediction confidence index. When the prediction confidence is lower than the preset threshold of 0.60, it issues control instructions to the active detection unit and the passive sniffing unit to increase the acquisition frequency. When the prediction confidence is higher than the preset threshold of 0.85, it issues control instructions to the active detection unit and the passive sniffing unit to decrease the acquisition frequency to reduce network overhead.

[0046] Please refer to Figure 3As shown, the data governance module includes a storage management module, a version control module, and a data service module. The storage management module includes a time-series database unit and a compression and archiving unit. The input of the time-series database unit establishes a data connection with the output of the data preprocessing module on the physical network side. The time-series database unit stores standardized state vectors collected from the physical network side according to time series. Each data record contains a timestamp, node identifier, and a complete K-dimensional feature vector. The time-series database unit uses a distributed time-series database engine to support high-concurrency writes and range queries. The input of the compression and archiving unit establishes a data connection with the output of the time-series database unit. The compression and archiving unit dynamically adjusts the sampling compression rate of historical data based on the monitoring results issued by the resource scheduling unit of the synchronization monitoring module. When the storage I / O load exceeds a preset threshold, the compression and archiving unit increases the compression rate to reduce the storage space occupied by the data. When the data query response time exceeds a preset threshold, the compression and archiving unit decreases the compression rate to improve the data reading speed. The version control module establishes a bidirectional data connection with the storage management module. The version control module trains the digital twin model... The training dataset is version-managed. Each time the twin network initiates a model update, the version control module reads the current complete dataset from the storage management module and creates a new data version snapshot. The version control module retains the 10 most recent version snapshots for rollback and comparison. When the collaborative evaluation end detects a continuous increase in model prediction bias, the version control module responds to the rollback command and restores the dataset to the previous version. The data service module includes a standardized interface unit and a permission management unit. The input end of the standardized interface unit establishes a data connection with the output end of the storage management module, and the output end of the standardized interface unit establishes a data connection with the topology twin module and behavior twin module of the twin network. The standardized interface unit provides a unified data query API to the twin network, which supports data retrieval by time range parameters, node range parameters, and feature dimension parameters. The permission management unit establishes a control connection with the standardized interface unit. The permission management unit assigns different access permission levels to different data users, controlling the read and write permissions of each data user to different node ranges and different feature dimensions in the dataset.

[0047] Please refer to Figure 4As shown, the twin network includes a topology twin module, a behavior twin module, and a prediction module. The topology twin module includes a topology discovery unit and a topology update unit. The input of the topology discovery unit establishes a data connection with the output of the standardized interface unit of the data governance end. The topology discovery unit identifies the connection relationships between nodes in the physical network through a network layer topology discovery protocol, which includes link layer discovery protocol and simple network management protocol topology discovery. The topology discovery unit generates a network topology graph containing a set of nodes and a set of edges from the identified nodes and connection relationships. This network topology graph is an undirected graph or a directed graph data structure. The input of the topology update unit establishes an indirect data connection with the output of the feedback execution module of the physical network end through the data governance end. According to the connection, the topology update unit detects topology changes based on real-time status data from the physical network. When a change in link connectivity or routing is detected, the topology update unit triggers an incremental update of the topology graph. The incremental update only modifies the changed nodes or edges without reconstructing the entire topology graph. The behavioral twin module includes a traffic modeling unit and a channel modeling unit. The input of the traffic modeling unit establishes a data connection with the output of the standardized interface unit of the data governance end. The traffic modeling unit builds a traffic arrival model based on historical traffic data provided by the data governance end. This traffic arrival model adopts a self-similar traffic model or a Markov-modulated Poisson process model. The traffic modeling unit estimates the model parameters from the historical traffic data and stores the estimated model parameters in the behavioral twin. The module's local cache is used to generate traffic arrival patterns in simulations. The input of the channel modeling unit establishes a data connection with the output of the standardized interface unit of the data governance unit. The channel modeling unit builds a channel quality model, which calculates link-level channel attenuation and bit error rate based on signal strength, interference level, and signal-to-noise ratio parameters collected from the physical network. The channel modeling unit uses a statistical channel model based on Rayleigh or Ricean fading and performs online calibration of the model parameters using measured data. The input of the prediction module establishes data connections with the outputs of the topology twin module, the behavioral twin module, and the standardized interface unit of the data governance unit, respectively. The prediction module uses the standardized state vector at the current moment as the initial... The initial condition input is a digital twin model, which is composed of a network topology map provided by the topology twin module, a traffic arrival model provided by the behavior twin module, and a channel quality model. The inference and prediction module uses discrete event simulation or network calculus-based analysis methods for simulation and inference. In discrete event simulation mode, the inference and prediction module processes data packet sending, queuing, forwarding, and receiving events in the network according to the timestamp order. In network calculus mode, the inference and prediction module calculates the upper bound of end-to-end delay and the upper bound of backlog based on the arrival curve and the service curve. The inference and prediction module outputs a predicted state vector sequence for the next T time windows. This predicted state vector sequence includes the predicted bandwidth utilization, predicted end-to-end delay, predicted packet loss rate, and predicted jitter value for each link.The prediction module employs a multi-timescale prediction strategy. On a short timescale, it uses a deterministic analysis method based on network calculus to predict communication quality within 1 to 5 time windows, each window being 100ms long. On a long timescale, it uses a time-series prediction model trained on historical data to predict communication quality within 6 to 20 time windows, each window being 1 second long. This time-series prediction model uses a Long Short-Term Memory (LSTM) network or a gated recurrent unit (GRU) neural network architecture. The prediction module generates the final prediction value by weighted fusion of the short-timescale and long-timescale prediction results. The fusion weights are dynamically adjusted based on the stability of the current network state. The prediction module calculates the variance of the state vector over the most recent 10 time points as the network state fluctuation value. When this fluctuation value exceeds a preset threshold, the network state is considered to be highly volatile, and the weight of the short-timescale prediction is increased to above 0.7. When the fluctuation value is below the preset threshold, the network state is considered stable, and the weight of the long-timescale prediction is increased to above 0.7.

[0048] Please refer to Figure 5 As shown, the collaborative evaluation module includes a correlation calculation module, a parameter adaptation module, and a confidence generation module. The input of the correlation calculation module establishes data connections with the output of the feedback execution module of the physical network and the output of the inference prediction module of the twin network. After each prediction cycle, the correlation calculation module reads the actual state vector s_real collected by the physical network at time t and the state vector s_pred predicted by the twin network at time t-1. The correlation calculation module compares the two and calculates the quality correlation coefficient Φ_t. The formula for calculating this quality correlation coefficient is:

[0049] in, for The mass correlation coefficient at time t, its value range is: ; The dimension of the state vector; The first data actually collected at the physical network end Dimensional state value; The k-th dimension state value predicted by the twin network endpoint; For the first The weighting coefficients of the dimensional state variables are preset according to the importance of each dimension to communication quality; This is the similarity function.

[0050] After calculating the quality correlation coefficient, the correlation calculation module will... Compared with a preset threshold, when When the accuracy exceeds the preset threshold of 0.85, the correlation calculation module determines that the current prediction of the twin network has high accuracy and outputs a high accuracy flag to the parameter adaptation module. When the error falls below a preset threshold of 0.60, the correlation calculation module determines that the prediction deviation of the Siamese network is too large and outputs a deviation flag to the parameter adaptation module to trigger the parameter adjustment mechanism. The input of the parameter adaptation module establishes a data connection with the output of the correlation calculation module, and the output of the parameter adaptation module establishes data connections with both the inference and prediction module of the Siamese network and the feedback execution module of the physical network. The parameter adaptation module dynamically adjusts the simulation inference parameters of the Siamese network based on the quality correlation coefficient. When the value is below 0.60, the parameter adaptation module sends a command to the inference and prediction module to shorten the prediction time window. This command reduces the simulation time granularity from 10ms to 1ms in discrete event simulations by 50% of the original length and increases the granularity parameter of the simulation. When the value is higher than 0.90, the parameter adaptation module issues an instruction to the inference and prediction module to extend the prediction time window length T to 150% of the original length and reduce the simulation inference granularity parameter, that is, to increase the simulation time granularity from 10ms to 50ms to reduce computational overhead. Between 0.60 and 0.90, the parameter adaptation module keeps the current parameters unchanged and issues a command to the inference and prediction module to maintain the current simulation configuration; the input of the confidence generation module establishes a data connection with the output of the correlation calculation module. The confidence generation module reads the time series of the quality correlation coefficient from the correlation calculation module and generates the prediction confidence index C_t. The calculation formula for this prediction confidence index is:

[0051] in Take 0.50, Take 0.35, Take 0.15, Take 10, The confidence index is calculated by the normalized standard deviation of the quality correlation coefficient over the most recent 10 time points, which is the standard deviation divided by the difference between the maximum and minimum standard deviations. The feedback is sent to the physical network execution module and the policy parsing module at the policy execution end, respectively.

[0052] Please refer to Figure 6As shown, the policy execution end includes a policy parsing module, an instruction conversion module, and an effect verification module. The input of the policy parsing module establishes a data connection with the output of the confidence generation module of the collaborative evaluation end. The policy parsing module receives the network communication quality prediction value and prediction confidence index output by the collaborative evaluation end. Internally, the policy parsing module stores a preset policy rule base containing several mapping rules. Each mapping rule includes triggering conditions and a corresponding execution policy. Triggering conditions include predicted bandwidth utilization exceeding a preset threshold, predicted end-to-end latency exceeding a preset threshold, and predicted packet loss rate exceeding a preset threshold. Execution policies include traffic shaping policies, load balancing policies, and priority queue scheduling policies. The policy parsing module compares the received prediction value with each triggering condition in the policy rule base and matches the corresponding execution policy. When multiple rules are triggered simultaneously, the policy parsing module selects the execution policy with the highest priority according to a preset priority order. The input of the instruction conversion module establishes a data connection with the output of the policy parsing module. The instruction conversion module converts the matched execution policy into standard configuration instructions or control signaling for specific network devices. These standard configuration instructions include routing adjustment instructions, i.e., modifying network devices. The next-hop address in the routing table, bandwidth limiting commands (setting traffic policing or traffic shaping parameters on the network device interface), and queue scheduling commands (modifying the queue scheduling algorithm and weight parameters of the network device's outgoing ports) are all part of the command conversion module. The module has built-in command line templates for different device manufacturers. Based on the target network device's model and operating system version, the module selects the appropriate template and fills in the policy parameters to generate executable device configuration commands. The output of the command conversion module establishes a control connection with each network device in the physical network. The module then sends the generated device configuration commands to the target network device via Secure Shell Protocol or Simple Network Management Protocol. The input of the effect verification module establishes a data connection with the output of the latency monitoring unit of the synchronization monitoring module. After the command conversion module issues the execution action, the effect verification module obtains the network status change data after execution through the synchronization monitoring module. The module compares the actual change effect with the predicted change effect, calculates the relative deviation between the actual and predicted values, and generates a policy execution effect report containing the execution policy identifier, issuance time, actual change value, and relative deviation. This report is then fed back to the parameter adaptation module at the collaborative evaluation end.

[0053] Please refer to Figure 7As shown, the synchronous monitoring module includes a delay monitoring unit, a resource scheduling unit, and an anomaly early warning unit. The input of the delay monitoring unit establishes monitoring data connections with the feedback execution module of the physical network, the storage management module of the data governance end, the simulation prediction module of the twin network, and the instruction conversion module of the policy execution end, respectively. The delay monitoring unit collects data acquisition delays from the physical network (i.e., the time interval from the probe message being sent to the data preprocessing completion), storage I / O delays from the data governance end (i.e., the time interval from data writing to the time-series database), simulation prediction time from the twin network end (i.e., the time interval from the state vector input to the predicted value output), and instruction issuance delays from the policy execution end (i.e., the time interval from the state vector input to the predicted value output). From the time interval between policy matching and device configuration taking effect, the latency monitoring unit aggregates the collected latency data from each end into a latency state vector and outputs it to the resource scheduling unit and the anomaly warning unit respectively. The input end of the resource scheduling unit establishes a data connection with the output end of the latency monitoring unit, and the output end of the resource scheduling unit establishes control connections with the simulation prediction module at the twin network end and the compression and archiving unit at the data governance end. The resource scheduling unit compares the simulation simulation time with a preset threshold of 500ms. When the simulation simulation time exceeds 500ms, the resource scheduling unit triggers a model simplification mechanism. This model simplification mechanism includes issuing an instruction to the simulation prediction module to reduce the simulation spatial resolution, i.e., reducing the amount of data in the simulation network. The resource scheduling unit attempts three simplification measures sequentially according to a preset simplification priority order until the simulation time drops below 500ms. These measures include: increasing the number of nodes involved in the calculation; issuing instructions to the simulation prediction module to reduce the simulation time granularity (i.e., increasing the minimum time unit for discrete event simulation); or issuing instructions to the simulation prediction module to enable a fast simulation mode (i.e., using network-based calculus analysis methods to replace discrete event simulation). The resource scheduling unit compares the storage I / O latency with a preset threshold of 200ms. When the storage I / O latency exceeds 200ms, the resource scheduling unit issues instructions to the compression and archiving unit at the data governance end to increase the compression ratio. The input of the anomaly warning unit is connected to the latency monitoring unit. The output end of the system establishes a data connection with the output end of the correlation calculation module. The anomaly warning unit receives the quality correlation coefficient sequence from the correlation calculation module. When the quality correlation coefficient is lower than the preset threshold of 0.50 for three consecutive periods, the anomaly warning unit sends an early warning signal to the network management system. The early warning signal includes the number of consecutive periods below the threshold and the latest quality correlation coefficient value. The early warning signal also includes text information indicating that there is a significant deviation between the digital twin model and the physical network and suggesting that the model be recalibrated. At the same time, the anomaly warning unit receives latency data from each end from the latency monitoring unit. When the latency of any end exceeds its respective preset severity threshold, the anomaly warning unit sends a corresponding latency anomaly warning signal to the network management system.

[0054] The following details the operating principle and process of the entire real-time prediction system for network communication quality based on digital twins: After system startup, initialization is performed first. The active detection unit in the data acquisition module of the physical network sends TWAMP probe messages to each target node in the physical communication network according to the preset initial detection frequency. The passive sniffing unit configures port mirroring policies on the core link and backbone link and begins to receive mirrored traffic. The data acquisition module continuously transmits the raw data collected, such as link bandwidth utilization, end-to-end latency, packet loss rate, jitter value, and signal strength, to the data preprocessing module. The data cleaning unit of the data preprocessing module performs outlier detection on the received raw data. When a certain data is detected... When a data point deviates from the mean by more than three standard deviations, the data cleaning unit marks it as an outlier and removes it. Then, the data cleaning unit fills the remaining gaps with the weighted average of data from the preceding and following time points using linear interpolation. The data cleaning unit also aligns data from the active detection and passive sniffing units at different sampling frequencies according to a unified timestamp, ensuring that all features at the same time are aligned on the time axis. The cleaned and aligned data is then passed to the feature extraction unit, which extracts the bandwidth utilization sequence, latency sequence, packet loss rate sequence, and jitter value sequence for each link. The feature extraction unit then concatenates these sequences according to a preset dimensional order to form a dimension. The standardized state vectors are stored in the time-series database unit via the data governance terminal and forwarded to the inference and prediction module of the twin network terminal. Simultaneously, the topology discovery unit of the topology twin module reads the connection state information of each node in the physical network through the data governance terminal and generates an initial network topology map. Each node in this topology map corresponds to a router or switch in the physical network, and each edge corresponds to a physical link or logical link. The traffic modeling unit of the behavior twin module reads the historical traffic data of the most recent 24 hours stored in the data governance terminal, estimates the Hurst parameter and Markov chain transition probability matrix from this historical traffic data to establish a traffic arrival model, and the channel modeling unit reads the historical data of signal strength, interference level, and signal-to-noise ratio of each link from the data governance terminal and builds a model accordingly. Establish channel quality models for each link; after receiving the standardized state vector at the current moment, the prediction module uses the bandwidth utilization of each link in the state vector as the initial traffic load condition, and the delay and jitter values ​​as the initial channel conditions. Combining the network topology map and routing table information provided by the topology twin module, and the traffic arrival model and channel quality model provided by the behavioral twin module, the prediction module performs discrete event simulation in the digital twin model. The prediction module advances the simulation timeline from the current moment forward, processing the data packets generated by each node, the queuing and waiting events on each link, and the data packet transmission events within each 100ms time window. After processing each time window, the prediction module records the bandwidth utilization, end-to-end delay, packet loss rate, and jitter value of each link at the end of the window, until all are completed. Simulation and output of time windows. The prediction module generates a sequence of predicted state vectors. For the first five windows of the short timescale, it independently calculates the end-to-end delay upper bound using network calculus and then weights and fuses this upper bound with the results of discrete event simulation. For the 6th to 20th windows of the long timescale, it uses a Long Short-Term Memory (LSTM) network model for time-series prediction and weights and fuses this prediction result with the discrete event simulation result. Finally, it generates a comprehensive predicted value sequence covering both short and long timescales. The prediction module outputs the generated predicted value sequence to the correlation calculation module at the collaborative evaluation end. At the end of each prediction cycle, the correlation calculation module reads the actual data collected during that cycle from the feedback execution module at the physical network end. The time-state vector is read from the inference and prediction module. Time prediction The time-state vector is used for correlation calculation. The module calculates the similarity between the actual vector and the predicted vector dimension by dimension, and then sums the similarities of each dimension according to preset weight coefficients and divides by the sum of the weight coefficients to obtain the quality correlation coefficient. ,when When the value is greater than 0.85, the correlation calculation module outputs a high-accuracy prediction indicator to the parameter adaptation module. When the correlation calculation module is below 0.60, it outputs a large prediction deviation flag to the parameter adaptation module. Upon receiving the flag, the parameter adaptation module executes two feedback paths simultaneously. In the first feedback path, the parameter adaptation module sends an instruction to the inference prediction module to shorten the prediction time window to 50% of its original length and increase the simulation inference granularity parameter. Upon receiving this instruction, the inference prediction module halves the T value of subsequent prediction periods and reduces the time granularity of discrete event simulation from 10ms to 1ms to improve prediction accuracy. In the second feedback path, the parameter adaptation module sends an instruction to the feedback execution module at the physical network end through the data governance end to increase the data acquisition frequency. Upon receiving this instruction, the feedback execution module sends a control signal to the active detection unit to increase the detection frequency to twice the original frequency and increase the number of detection target nodes by 50%. The active detection unit adjusts its detection behavior accordingly so that the physical network end can provide denser and more comprehensive real-time status data in subsequent prediction periods. The confidence generation module continuously reads the quality correlation coefficients of the last 10 moments from the correlation calculation module and calculates the prediction confidence index for the current moment. ,Should The data is simultaneously sent to the feedback execution module at the physical network end and the policy parsing module at the policy execution end. After receiving the prediction confidence index and the prediction state vector sequence, the policy parsing module compares the predicted bandwidth utilization, predicted end-to-end latency, and predicted packet loss rate of each link in the prediction state vector sequence with the trigger conditions in the policy rule base. When the predicted end-to-end latency exceeds a preset threshold of 100ms, the policy parsing module matches a priority queue scheduling policy; when the predicted bandwidth utilization exceeds a preset threshold of 80%, the policy parsing module matches a load balancing policy; and when the predicted packet loss rate exceeds a preset threshold of 1%, the policy parsing module matches a traffic shaping policy. The policy parsing module then outputs the matched policy identifier. The command is then sent to the command conversion module. Based on the target network device's model, the module selects the corresponding template from its built-in command-line template library, fills in the policy parameters, and generates a complete device configuration command. The command conversion module then sends the configuration command to the target network device via a secure shell protocol session and receives configuration confirmation information from the device. In the next prediction period after the command is sent, the effect verification module obtains the actual network status change data for that period through the synchronous monitoring module. The effect verification module compares the actual change with the predicted change and calculates the relative deviation. If the relative deviation is within the allowable range, the effect verification module sends a success report to the parameter adaptation module; if the relative deviation exceeds the allowable range... The effect verification module will then execute a deviation report and feed it back to the parameter adaptation module, which will trigger a retry mechanism for strategy adjustment. During system operation, the latency monitoring unit of the synchronization monitoring module continuously collects runtime latency data from each terminal at 100ms intervals and sends the latency data to the resource scheduling unit and the anomaly warning unit. When the resource scheduling unit detects that the simulation time exceeds 500ms, it will attempt to reduce the simulation spatial resolution and simulation time granularity sequentially according to a preset order until the simulation time drops below 500ms. When the anomaly warning unit detects that the quality correlation coefficient is below 0.50 for three consecutive periods, it will send a model deviation warning signal to the network management system along with the most recent... The three-cycle quality relational values ​​and corresponding network status snapshots are used. After receiving the early warning signal, the network management system notifies the network operation and maintenance personnel to recalibrate the digital twin model. When the connectivity of links in the physical network changes or the forwarding path changes due to routing protocol convergence, the topology update unit detects the topology change from the real-time status data of the physical network and triggers an incremental update of the topology graph. The topology update unit only deletes disconnected edges or adds newly established edges without retracing all nodes and edges in the network. This incremental update is completed within 100ms and the updated topology graph is pushed to the inference and prediction module for use in the next prediction cycle. When the prediction confidence index at the physical network end remains above 0.When the frequency exceeds 10 cycles, the feedback execution module sends a control signal to the active detection unit to reduce the detection frequency to 50% of the initial frequency and decrease the number of target nodes by 30%, in order to reduce the additional traffic load on the physical network caused by active detection. With the coordinated operation of all these components, the system continues to run, forming a complete closed loop from data acquisition to simulation and deduction, then to deviation detection and adaptive parameter adjustment, ensuring that the prediction accuracy of the digital twin model remains within a preset reliable range throughout long-term operation.

[0055] For the contents not described in detail, such as the specific message format of the TWAMP protocol, the specific training method of the Long Short-Term Memory network, and the specific implementation mechanism of event scheduling in discrete event simulation, they are all existing technologies that can be known or implemented by those skilled in the art based on common knowledge and conventional experimental methods; the specific programming language and implementation framework of each software module are not limitations of this invention, and any conventional software development method that can achieve the above functions can be used as an alternative.

[0056] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A real-time prediction system for network communication quality based on digital twins, characterized in that, It includes a physical network terminal, a data governance terminal, a twin network terminal, a collaborative evaluation terminal, and a policy execution terminal, all of which are equipped with a synchronous monitoring module. The physical network terminal is used to collect the operating status data of each network node in the physical communication network in real time, and convert the collected raw data into a standardized state vector through data preprocessing; the physical network terminal is also used to receive quality prediction feedback from the collaborative evaluation terminal, and adjust the data collection frequency and collection granularity according to the feedback information. The twin network terminal is used to construct a digital twin model of the physical communication network based on the real-time status data and historical operation data of the physical network terminal, and to simulate and deduce by inputting the physical network state vector at the current moment into the digital twin model, and output the predicted value of network communication quality for multiple future time windows. The collaborative evaluation terminal is used to compare the state vector actually collected by the physical network terminal with the state vector predicted by the twin network terminal after each prediction cycle, calculate the quality correlation coefficient used to characterize the prediction accuracy, and issue collection strategy adjustment instructions to the physical network terminal and simulation inference parameter adjustment instructions to the twin network terminal according to the quality correlation coefficient, forming a closed-loop collaborative mechanism from data collection at the physical network terminal to simulation inference at the twin network terminal and then to deviation detection and feedback adjustment at the collaborative evaluation terminal. The synchronous monitoring module is used to monitor in real time the data acquisition delay at the physical network end, the simulation and deduction time at the twin network end, and the changing trend of the quality correlation coefficient at the collaborative evaluation end. When the data acquisition delay exceeds a preset threshold, the acquisition strategy is adjusted. When the simulation and deduction time exceeds a preset threshold, the model simplification mechanism is triggered. When the quality correlation coefficient is continuously lower than a preset threshold, an early warning signal is issued.

2. The system according to claim 1, characterized in that, The operational status data of the physical network terminal includes link bandwidth utilization, end-to-end latency, packet loss rate, jitter value, and signal strength; the physical network terminal includes a data acquisition module, a data preprocessing module, and a feedback execution module; The data acquisition module includes an active detection unit and a passive sniffing unit. The active detection unit is used to send probe messages to target nodes in the network and measure end-to-end latency and packet loss rate by measuring the round-trip time and packet loss of the probe messages. The passive sniffing unit is used to obtain real-time traffic data on the network link through port mirroring or traffic mirroring and to statistically analyze the link bandwidth utilization and traffic distribution characteristics. The data preprocessing module includes a data cleaning unit and a feature extraction unit. The data cleaning unit performs outlier detection and removal, missing value imputation, and time alignment operations on the collected raw data. The feature extraction unit extracts network communication quality feature vectors from the cleaned data. The feature vectors include bandwidth utilization sequences, latency sequences, packet loss rate sequences, and jitter value sequences for each link, and concatenates these sequences to form a standardized state vector. The feedback execution module receives quality prediction feedback from the collaborative evaluation terminal and dynamically adjusts the data collection frequency based on the prediction confidence index in the feedback: when the prediction confidence is lower than a preset threshold, the data collection frequency and probe packet sending density are increased; when the prediction confidence is higher than the preset threshold, the data collection frequency is reduced to decrease network overhead.

3. The system according to claim 1, characterized in that, The data governance terminal is used to uniformly store, manage versions, and verify the quality of multi-source heterogeneous operational status data collected by the physical network terminal, construct a time-series dataset containing historical and real-time status data, and provide a standardized data access interface for the twin network terminal; the data governance terminal is also used to dynamically adjust the retention strategy and sampling compression rate of historical data based on the monitoring results of the synchronization monitoring module. The data governance module includes a storage management module, a version control module, and a data service module. The storage management module includes a time-series database unit and a compression and archiving unit. The time-series database unit stores standardized state vectors collected from the physical network in a time-series manner, with each data record containing a timestamp, node identifier, and complete feature vector. The compression and archiving unit dynamically adjusts the sampling compression rate of historical data based on the monitoring results of the synchronization monitoring module. When the storage I / O load exceeds a preset threshold, the compression rate is increased; when the data query response time exceeds a preset threshold, the compression rate is decreased. The version control module is used to manage the training dataset of the digital twin model. Each time the twin network initiates a model update, a new data version snapshot is created, and the most recent preset number of versions are retained for rollback and comparison. The data service module includes a standardized interface unit and a permission management unit. The standardized interface unit is used to provide a unified data query API to the twin network, supporting data retrieval by time range, node range, and feature dimension. The permission management unit is used to control the access permissions of different data users to the dataset.

4. The system according to claim 1, characterized in that, The twin network terminal includes a topology twin module, a behavior twin module, and an inference and prediction module; The topology twin module includes a topology discovery unit and a topology update unit. The topology discovery unit is used to identify the connection relationships between nodes in the physical network through a network layer topology discovery protocol, generating a network topology graph containing a set of nodes and a set of edges. The topology update unit is used to detect topology changes based on real-time status data from the physical network. When a change in link connectivity or routing is detected, an incremental update of the topology graph is triggered. The behavioral twin module includes a traffic modeling unit and a channel modeling unit. The traffic modeling unit is used to establish a traffic arrival model based on historical traffic data provided by the data governance terminal. The traffic arrival model adopts a self-similar traffic model. Alternatively, a Markov-modulated Poisson process model can be used to generate traffic arrival patterns in simulations; the channel modeling unit is used to establish a channel quality model, which calculates link-level channel attenuation and bit error rate based on signal strength, interference level, and signal-to-noise ratio parameters; the prediction module is used to input the standardized state vector at the current moment as initial conditions into the digital twin model, and perform simulations using discrete event simulation or network calculus-based analysis methods, outputting a predicted state vector sequence for the next T time windows, which includes the predicted bandwidth utilization, predicted end-to-end delay, predicted packet loss rate, and predicted jitter value for each link.

5. The system according to claim 4, characterized in that, The inference and prediction module adopts a multi-timescale prediction strategy: on a short timescale, a deterministic analysis method based on network computation is used to predict the communication quality within 1 to 5 time windows in the future, with each time window being 100ms long; on a long timescale, a time series prediction model trained based on historical data is used to predict the communication quality within 6 to 20 time windows in the future, with each time window being 1s long. The short-timescale prediction results and long-timescale prediction results are weighted and fused to generate the final prediction value. The fusion weight is dynamically adjusted according to the stability of the current network state. When the network state fluctuates greatly, the weight of the short-timescale prediction is increased.

6. The system according to claim 1, characterized in that, The correlation calculation module is used to process the actual data collected at the physical network end after each prediction period. Time-state vector and Siamese network endpoints Time prediction By comparing the state vectors at different times, the quality correlation coefficient is calculated using the following formula: in, for The mass correlation coefficient at time t, its value range is: ; is the dimension of the state vector; The first data actually collected at the physical network end Dimensional state value; The k-th dimension state value predicted by the twin network endpoint; For the first The weighting coefficients of the dimensional state variables are preset according to the importance of each dimension to communication quality; For similarity functions; when the quality correlation coefficient When the value is greater than the preset threshold of 0.85, the current prediction of the Siamese network is considered to have high accuracy; when... When the value is below the preset threshold of 0.60, it is determined that the prediction deviation of the twin network is too large, triggering the parameter adjustment mechanism; The parameter adaptive module is used to dynamically adjust the simulation and inference parameters of the Siamese network based on the quality correlation coefficient, specifically including: when When the value is below 0.60, shorten the prediction time window length. And increase the precision parameters of the simulation; when When the value is higher than 0.90, the prediction time window length T is increased and the simulation refinement parameter is reduced to decrease computational overhead; when When the value is between 0.60 and 0.90, the current parameter remains unchanged; the confidence generation module is used to generate a predicted confidence index based on the time series of the quality correlation coefficient, as shown in the following formula: in, for The prediction confidence index at time t, with a value range of . ; The quality correlation coefficient at the current moment; For the most recent The moving average of the mass correlation coefficient at each time point; For the most recent Normalized standard deviation of the quality correlation coefficient at each time point; For the weighting coefficients, satisfying .

7. The system according to claim 1, characterized in that, The policy execution terminal receives the predicted network communication quality value and prediction confidence index output by the collaborative evaluation terminal, converts the prediction result into network scheduling instructions or traffic optimization actions according to preset policy rules, and feeds back the execution result to the synchronous monitoring module to verify the policy effect. The policy execution terminal includes a policy parsing module, an instruction conversion module, and an effect verification module. The policy parsing module receives the predicted network communication quality value and prediction confidence index output by the collaborative evaluation terminal, matches the corresponding execution policy according to a preset policy rule base, which contains several mapping rules. The instruction conversion module converts the matched execution policy into standard configuration instructions or control signaling for specific network devices. The standard configuration instructions include routing adjustment instructions, bandwidth limiting instructions, and queue scheduling instructions. The effect verification module, after the instruction conversion module issues the execution action, obtains the network status change data after execution through the synchronous monitoring module, compares the actual change effect with the predicted change effect, generates a policy execution effect report, and feeds it back to the collaborative evaluation terminal.

8. The system according to claim 1, characterized in that, The synchronization monitoring module includes a latency monitoring unit, a resource scheduling unit, and an anomaly warning unit. The latency monitoring unit collects data acquisition latency at the physical network end, storage I / O latency at the data governance end, simulation and deduction time at the twin network end, and instruction issuance latency at the policy execution end every preset period. The resource scheduling unit triggers a model simplification mechanism when the simulation and deduction time exceeds a preset threshold. This mechanism includes reducing the simulation spatial resolution, reducing the simulation time granularity, or enabling a fast simulation mode. When the storage I / O latency exceeds a preset threshold, it triggers an adjustment to the compression and archiving strategy at the data governance end. The anomaly warning unit sends a warning signal to the network management system when the quality correlation coefficient is below a preset threshold of 0.50 for three consecutive periods, indicating a significant deviation between the digital twin model and the physical network, and recommends model recalibration.

9. The system according to claim 2, characterized in that, The active detection unit sends detection messages using the UDP-based TWAMP protocol or ICMP timestamp request messages. The active detection unit dynamically adjusts the sending frequency of detection messages and the number of target nodes based on the prediction confidence fed back by the collaborative evaluation end. When the prediction confidence is lower than 0.60, the detection frequency is increased to twice the original frequency, and the number of target nodes is increased by 50%.

10. A method for real-time prediction of network communication quality based on digital twins, applicable to the real-time prediction system for network communication quality based on digital twins according to any one of claims 1-9, characterized in that, Includes the following steps: S1: The data acquisition module at the physical network end collects the operating status data of each network node in the physical communication network through active detection and passive sniffing. The data preprocessing module converts the collected raw data into a standardized state vector. S2: The data governance end performs unified storage and version management of the multi-source heterogeneous operation status data collected by the physical network end, and constructs a time-series dataset containing historical status data and real-time status data to provide a standardized data access interface for the twin network end. S3: The topology twin module constructs and continuously updates the network topology map based on the real-time status data provided by the data governance terminal, and the behavior twin module establishes the traffic arrival model and the channel quality model based on historical traffic data and channel data. S4: The simulation and prediction module takes the standardized state vector at the current moment as the initial condition input to the digital twin model, performs simulation and deduction through discrete event simulation or network calculation analysis methods, and outputs the predicted value of network communication quality for multiple future time windows. S5: At the end of each prediction cycle, the correlation calculation module of the collaborative evaluation end compares the actual state vector collected by the physical network end with the predicted state vector by the twin network end, and calculates the quality correlation coefficient. ; S6: The parameter adaptive module is based on the quality correlation coefficient. The prediction time window length and simulation granularity parameters of the twin network are dynamically adjusted. The confidence generation module generates a prediction confidence index based on the time series of the quality correlation coefficient and sends an acquisition strategy adjustment instruction to the physical network. S7: The policy execution end receives the network communication quality prediction value and prediction confidence index output by the collaborative evaluation end, and converts the prediction results into network scheduling instructions or traffic optimization actions according to the preset policy rules and issues them for execution. S8: The synchronous monitoring module monitors the running latency and performance indicators of each terminal in real time. When the latency exceeds the preset threshold, the corresponding adjustment mechanism is triggered. When the quality correlation coefficient is continuously lower than the preset threshold, an early warning signal is issued.

Images