Predictive remote communication method and system

Abstract

The invention relates to the field of remote communication network optimization, and discloses a predictive remote communication method and system. The method comprises the following steps: generating a time sequence sample structure based on historical communication data, and constructing a deployable model through LSTM model training and Bayesian optimization; receiving a communication request in real time, predicting a demand probability, and triggering a dynamic handshake protocol to be connected with the security encryption; generating optimal path configuration by combining network topology analysis and a dynamic programming algorithm; and finally, online updating of the model is realized through performance monitoring and incremental learning. According to the method, time sequence prediction and dynamic path optimization are fused, the problems that model training and network topology are disjointed, dynamic optimization lags and the like in a traditional method are solved, the connection establishment speed and the transmission stability in a high-delay scene are remarkably improved, and meanwhile the self-adaptive capacity of the system is enhanced through embedded packaging and a hot updating mechanism.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of remote communication network optimization, and in particular to a method and system for predicting remote communication. BACKGROUND

[0002] In the field of remote communication network optimization, the existing solutions related to a method and system for predicting remote communication generally plan communication paths based on static rules or simple statistical models, which have limitations such as disconnection between model training and network topology analysis, lag in dynamic path optimization, separation of delay evaluation and connection prediction, etc. The existing methods mostly use a processing architecture that separates independent time series prediction from offline path calculation. In scenarios with high remote communication connection establishment delay, the historical communication patterns are not effectively mined, real-time network state changes are not responded to in a timely manner, and optimal path selection lacks prediction support, which makes it difficult to meet the stable implementation of fast establishment and stable transmission of communication connection in high delay sensitive scenarios. For the joint processing of historical communication data and network topology structure, the existing technology generally lacks a deep fusion mechanism of long short-term memory network model and dynamic programming algorithm, which makes it difficult to form a consistent process of data feature extraction-weight parameter optimization-path state evaluation-transmission configuration generation in complex network environment, resulting in the inability to realize pre-prediction and dynamic cooperative optimization of communication connection initiation time and transmission path selection. SUMMARY

[0003] The present application provides a method and system for predicting remote communication to solve the problem of how to train a long short-term memory network model and optimize a dynamic programming path based on historical communication data and network topology to realize early connection initiation and transmission path selection in scenarios with high remote communication connection establishment delay.

[0004] To solve the above technical problems, the present application provides a method for predicting remote communication, comprising: Obtain historical communication data, identify and clean abnormal values based on box plot or standard deviation, perform normalization processing using minimum-maximum normalization or Z-score standardization, and set the size and step of the sliding window according to the time granularity of the communication demand to perform sliding window segmentation, and generate time series sample structure; Obtain the time series sample structure, train a multi-layer LSTM model, determine the number of hidden layer units through an empirical formula, extract weight parameters, perform cross-validation optimization using Bayesian optimization, and perform model pruning and weight quantization to encapsulate the deployable model structure; Receive real-time communication requests, perform feature encoding and standardization processing when converting feature vectors, use the LSTM model to perform demand probability prediction, and set the size and weight coefficient of the sliding window to perform sliding average calculation to generate trigger instructions; According to the trigger instruction, a handshake protocol including connection initialization and identity authentication is initiated, bandwidth verification is performed based on real-time link load, and secure encryption processing is performed using AES and ECDSA to generate a verified connection; The verified connection is obtained, network topology analysis is performed based on real-time network state, dynamic programming calculation is performed using an improved Bellman equation, and delay evaluation processing is performed by segmenting and accumulating delays to generate an optimal path configuration; Performance indicators are collected during monitoring and transmission, error indicators including transmission delay deviation and packet loss rate fluctuation are extracted, incremental learning is performed, and online deployment processing is performed using a hot update mechanism to generate an updated model.

[0005] Further, the process of training the multi-layer LSTM model based on the time sequence sample structure includes: Data format verification and integrity detection are performed on the time sequence sample structure to exclude missing or abnormal samples; An input sequence of the long short-term memory network model is constructed, and a multi-layer LSTM structure is used to capture long-term dependencies and nonlinear dynamic changes in time series; Batch training is performed based on the time sequence sample structure, and the backpropagation algorithm is combined with the gradient descent optimization method to iteratively adjust network weight parameters; Early stopping mechanism and regularization techniques are introduced to prevent overfitting, and cross-validation is used to divide the training set and the validation set to dynamically monitor model performance indicators such as loss function value and prediction error.

[0006] Further, the process of extracting weight parameters and performing cross-validation optimization using Bayesian optimization includes: Based on the preset cross-validation strategy, the training data set is divided into multiple mutually exclusive subsets, one of which is used as the validation set and the rest as the training set, and the LSTM model is repeatedly trained; The optimization process includes adjusting network layer number, hidden unit number, learning rate, and regularization coefficient hyperparameters, and systematically exploring the parameter space through Bayesian optimization method; After each round of training, the performance indicators of the model on the validation set are calculated, and the optimization log is recorded for performance analysis; After completing all cross-validation rounds, the performance of the optimal weight parameter combination is selected as the final model parameter.

[0007] Further, the process of performing model pruning and weight quantization to encapsulate the deployable model structure includes: The model structure and parameters are converted to adapt to the computational resources and storage limitations of the embedded deployment environment, and model pruning and weight quantization techniques are used to reduce model size and computational complexity; The optimized model is packaged into an executable deployment unit, which includes model files, inference engine interfaces, and necessary runtime configuration files; During the encapsulation process, consistency verification is performed, including model inference accuracy detection and resource consumption assessment.

[0008] Furthermore, the process of receiving real-time communication requests and performing feature encoding and standardization when transforming feature vectors includes: Receive real-time communication requests initiated by remote communication terminals or application layers. These requests include timestamps, communication types, request frequencies, data packet sizes, terminal identifiers, and network status information. Preprocess the real-time communication request data, including data format validation, missing field completion, and outlier filtering; Valid communication request data is mapped to feature vectors. Feature encoding methods are used to convert categorical data into numerical vectors, and numerical data is standardized according to the model input requirements.

[0009] Furthermore, the process of performing demand probability forecasting using an LSTM model includes: The input features are analyzed in terms of time dimension to identify timestamp information and time series features contained in the feature vector; A time series input is constructed based on time dimension information and fed into an embedded long short-term memory network model; LSTM models capture long-term dependencies and short-term dynamic changes in time series through their gating mechanism, and progressively calculate the hidden state and cell state of the input sequence to achieve probabilistic prediction of future connection needs. The output of the LSTM model is post-processed, including probability normalization, outlier filtering, and confidence assessment.

[0010] Furthermore, the process of setting the sliding window size and weighting coefficients, performing a moving average calculation, and generating trigger instructions includes: The size and weighting coefficients of the moving average window are set. The window size is dynamically adjusted according to the time sensitivity of communication needs and the system response capability. The weighting coefficients adopt an exponential weighting method to enhance the influence of recent prediction results. Perform a moving average calculation on the continuous sequence of prediction results, and calculate the weighted average of the prediction probabilities within the current window point by point; The probability value after the moving average is compared with the preset trigger threshold, which is dynamically adjusted according to historical communication load and system resource status. Under conditions that are met or exceeded, a connection trigger command is automatically generated. The command includes the trigger time, expected connection type, priority, and related parameters.

[0011] Furthermore, a multi-layer LSTM model is trained. The number of hidden layer units is determined using an empirical formula. Weight parameters are extracted, and the expression for cross-validation tuning using Bayesian optimization includes: Network parameters are dynamically set based on the feature dimension, the input layer dimension is aligned with the feature vector length, and the number of hidden layer units is determined by an empirical formula. Confirmed, among which Define an improved cross-entropy loss function for the input feature dimension:

[0012] in, Improved cross-entropy loss function for LSTM training: This represents the batch sample size. Sample index; For the first The true label of each sample; For LSTM The hidden state of the time step; For output layer weights and biases; The regularization coefficient is used. Use the Sigmoid activation function; L2 norm squared; Number of hidden layer units in LSTM; As the optimization objective of backpropagation, it drives the adjustment of network weights and outputs the initial prediction model, which includes the network structure and weight parameters. Furthermore, weight parameters are extracted from the initial prediction model. With the hidden state matrix The training and validation sets are divided based on a cross-validation strategy. Bayesian optimization is used to fine-tune the cross-validation, specifically by dynamically adjusting hyperparameters using Bayesian optimization and defining a weighted performance metric for multi-fold cross-validation. in, The weighted performance metric for multi-fold cross-validation; For fold index: For cross-validation folds; For the first The weighting coefficient of the fold; Mean square error; These are the weighting coefficients for MAE; For the first The mean absolute error of the fold; Used to evaluate the generalization performance of different hyperparameter combinations, driving the grid search process, the optimized weight parameters are used to generate the final parameters through a weight fusion strategy:

[0013] in, These are the final weight parameters after fusion; For model indexing; The number of optimal parameter combinations; For the first The weights of each model; For the first The weight matrix of each model.

[0014] Furthermore, dynamic programming calculations are performed using the improved Bellman equation, and delay evaluation is carried out by piecewise delay accumulation. The expression for the optimal path configuration includes: The path weights are calculated using a weighted summation model, and a path score integrating multiple indicators is defined: in, Score the path weight; Link bandwidth; For time delay; Packet loss rate; These are the weighting coefficients; This is the normalized baseline value; Furthermore, The path weights are used as the basis for calculation and input into the dynamic programming algorithm. The improved Bellman equation is used for dynamic programming calculation to solve for the optimal path.

[0015] in, For nodes The minimum cumulative cost; For the current node; Adjacent nodes; for The set of adjacent nodes; For edge cost; Furthermore, the optimal path is selected through multi-objective ranking: in, Identify the optimal path; Candidate paths; For the candidate path set; For path Weighted scoring; Path stability score; As a weighting factor; The parameter corresponding to the maximum value.

[0016] Furthermore, a predictive remote communication system, applied to the method described in any of the above, includes: The node configuration module is used to obtain communication node configuration parameters and complete network topology constraints. The data acquisition module is used to collect real-time transmitted data packets and historical delay records; The feature processing module is used to extract features from real-time transmitted data packets and normalize historical delay records. The synchronization module is used to timestamp-align the transmitted feature vector with the standard delay sequence; The arbitration module is used to perform link congestion arbitration based on a dynamic threshold policy table; The control module is used to output path switching instructions to the routing controller; The update module is used to record the effective timestamp and update the dynamic threshold strategy table.

[0017] The key innovations of this invention include: (1) By acquiring the temporal sample structure, a Long Short-Term Memory (LSTM) network is trained to form an initial prediction model, thereby realizing the joint analysis and dynamic optimization of historical communication data and network topology.

[0018] (2) Extract weight parameters from the initial prediction model and perform cross-validation to generate an optimized model, ensuring deep integration of model training and network topology analysis.

[0019] (3) Embedded encapsulation of the optimized model to generate a deployable model structure to support the dynamic collaborative optimization of the timing of communication connection initiation and the selection of transmission path.

[0020] The following are its main beneficial effects: (1) By acquiring the temporal sample structure and training the LSTM network, the initial prediction model can effectively mine historical communication patterns and respond to real-time network status changes in scenarios with high latency in establishing remote communication connections, thus solving the problem that historical communication patterns have not been effectively mined in the existing technology.

[0021] (2) The optimized model is tuned by cross-validation of weight parameters to form a consistent process of model training and network topology analysis, ensuring that data feature extraction and transmission configuration generation are realized in complex network environments, and solving the problem of lagging dynamic path optimization in the existing technology.

[0022] (3) The deployable model structure supports the advance prediction of the timing of communication connection initiation and the dynamic collaborative optimization of transmission path selection through embedded encapsulation, ensuring that communication connection is quickly established and stably transmitted in high latency sensitive scenarios, and solving the problem of lack of predictive support for optimal path selection in existing technologies. Attached Figure Description

[0023] Figure 1 A flowchart illustrating a method for predicting remote communication provided in an embodiment of this application; Figure 2 This is a structural block diagram of a predictive remote communication system provided in an embodiment of this application. Detailed Implementation

[0024] Example 1: Refer to Figure 1 This is a flowchart illustrating a method for predicting remote communication provided in an embodiment of the present invention. The process may include at least steps S100-S600: S100. Acquire historical communication data, identify and clean outliers based on box plots or standard deviation, perform normalization processing using min-max normalization or Z-score normalization, and perform sliding window segmentation according to the time granularity of communication requirements to generate time series sample structure. S200. Obtain the temporal sample structure, train a multi-layer LSTM model, determine the number of hidden layer units through empirical formulas, extract weight parameters, perform cross-validation tuning using Bayesian optimization, and perform model pruning and weight quantization to encapsulate a deployable model structure. S300 receives real-time communication requests, performs feature encoding and standardization when transforming feature vectors, uses an LSTM model to perform demand probability prediction, sets the sliding window size and weight coefficients to perform moving average calculation, and generates trigger commands. S400 initiates a handshake protocol including connection initialization and authentication based on the trigger command, performs bandwidth verification based on real-time link load, and uses AES and ECDSA for security encryption to generate a verified connection. S500: After obtaining verification and connection, perform network topology analysis based on real-time network status, use the improved Bellman equation for dynamic programming calculation, and perform segmented delay accumulation for delay evaluation to generate the optimal path configuration. The S600 monitors the transmission process, collects performance indicators, extracts error indicators including transmission delay deviation and packet loss rate fluctuation, performs incremental learning, and uses a hot update mechanism for online deployment processing to generate an updated model.

[0025] Step S100 includes at least steps S110-S130: S110. Obtain historical communication data, perform outlier cleaning, and obtain the cleaned dataset. The steps for acquiring historical communication data, performing outlier cleaning, and obtaining a cleaned dataset specifically include the following: The historical communication data refers to the original communication records stored by the system, covering multi-dimensional information such as connection requests, transmission rates, latency, and packet loss rates during the communication period, serving as the input source for this step. First, outliers in the historical communication data are identified by setting anomaly detection rules. These outliers include, but are not limited to, extreme data points exceeding statistical thresholds, missing data, and data items with incorrect formats. Specifically, statistical analysis methods are used, such as boxplot anomaly detection or outlier identification based on standard deviation, combined with time series consistency checks, to filter out abnormal samples. Further, the identified outliers undergo cleaning processing, including outlier removal, missing value imputation, and format standardization. The imputation method employs linear interpolation based on nearby time points or a filling strategy based on historical similar patterns to ensure data continuity and integrity. During the cleaning process, the system automatically records a processing log, including the location of outlier data, processing method, and results, for subsequent auditing and performance analysis. After cleaning, the historical communication data is transformed into a structured cleaned dataset. Specifically, this dataset is stored in a standardized format, including timestamps, communication parameters, and their corresponding values, ensuring data consistency and traceability. This cleaned dataset serves as the output field name for this step and is subsequently called by the "Data to be Processed" section in S120, serving as the input basis for further feature extraction and normalization processing. Simultaneously, this dataset also provides high-quality data support for subsequent machine learning model training, permeating the overall communication demand prediction process of this invention.

[0026] S120. Extract time series features from the cleaned dataset, perform normalization processing, and generate preprocessed data; The steps of extracting time-series features from the cleaned dataset, performing normalization, and generating preprocessed data specifically include the following implementation process. The cleaned dataset serves as the input source. First, time-series features are extracted from it, covering key indicators such as communication request frequency, connection duration, latency trends, and bandwidth utilization. Specifically, a sliding window technique is used to segment the time-series data, calculating statistical features within each time window, such as mean, variance, maximum, minimum, and trend change rate, further extracting periodic and seasonal features. During feature extraction, considering the temporal characteristics of the communication scenario, the Autocorrelation Function (ACF) and Partial Autocorrelation Function (PACF) are used to assist in identifying effective time dependencies. After extraction, normalization is performed for different feature dimensions, specifically using Min-Max Normalization or Z-score standardization to map feature values ​​to a uniform numerical range or a zero-mean, unit-variance distribution to eliminate the impact of dimensional differences on subsequent model training. During normalization, the system dynamically adjusts parameters to adapt to changes in feature distribution, ensuring data stability and model generalization ability. After processing, the generated preprocessed data is stored in a structured format, containing multi-dimensional feature vectors corresponding to timestamps, which serve as the output field names for this step. This preprocessed data is subsequently consumed by the "input data" of S130 to construct a time-series sample structure, thereby supporting the training and optimization of machine learning models, which is integral to the core process of predicting remote communication needs.

[0027] S130. Perform sliding window segmentation on the preprocessed data to generate a time series sample structure; The steps of generating a time-series sample structure by performing sliding window segmentation on the preprocessed data specifically include the following implementation details. The preprocessed data serves as input data. First, the size and step size parameters of the sliding window are set. The window size is determined based on the time granularity of communication requirements, while the step size controls the degree of overlap between samples. Specifically, the system extracts data segments of consecutive time periods from the preprocessed data according to the parameters, forming individual time-series samples. Each time-series sample contains multi-dimensional normalized features, reflecting the dynamic changes in communication status within that time period. Furthermore, to enhance the model's capture of time dependencies, the sliding window employs an overlapping strategy, meaning that there is a certain proportion of data reuse between adjacent samples to ensure temporal continuity. The generated time-series sample structure is stored in a serialized format, including the sample start and end times, feature vector sequences, and corresponding labels (such as historical connection demand strength), satisfying the input data format requirements of Long Short-Term Memory (LSTM) networks. During sample generation, the system performs filtering or compensation for abnormal or missing data segments to ensure sample quality. The time-series sample structure serves as the output field name for this step and is subsequently called by the "training data" in S210 for the training process of the LSTM model, forming the basic data structure for predicting remote communication needs, which runs through the machine learning model training module of this invention.

[0028] Steps S110 to S130 systematically complete the transformation process from raw data to model training input through cleaning historical communication data, feature extraction, and time-series sample construction, forming a high-quality, structured training data foundation that supports efficient training and accurate prediction of subsequent machine learning models.

[0029] Step S200 includes at least steps S210-S230: S210. Obtain the temporal sample structure, perform LSTM network training to obtain the initial prediction model; The system obtains the time-series sample structure, specifically using it as the input source. This time-series sample structure, generated in the previous step S130, contains multi-dimensional normalized time-series features after sliding window segmentation and their corresponding historical connection requirement labels. First, the system performs data format verification and integrity checks on the time-series sample structure, excluding missing or abnormal samples to ensure the validity and consistency of the input data. Further, for the valid samples, an input sequence for a Long Short-Term Memory (LSTM) network model is constructed. Specifically, a multi-layer LSTM structure is used to capture long-term dependencies and nonlinear dynamic changes in the time series. The structural parameters of the LSTM network include the input layer dimension, the number of hidden layer units, the number of layers, and the output layer configuration, all dynamically set according to the feature dimensions of the historical communication data and the prediction requirements. Subsequently, the system performs batch training based on the time-series sample structure, using a backpropagation algorithm combined with gradient descent optimization to iteratively adjust the network weight parameters. During training, early stopping and regularization techniques are introduced to prevent overfitting. Simultaneously, cross-validation is used to divide the training and validation sets, and model performance metrics such as loss function values ​​and prediction errors are dynamically monitored. After training, the system outputs an initial prediction model, which includes the trained network structure and corresponding weight parameters. This model is designated as the output field name "Initial Prediction Model" for this step and passed to the "Model to be Optimized" in subsequent step S220 for further processing. This initial prediction model not only serves as the core product of the machine learning model training module of this invention but also provides the basic prediction capability for the subsequent connection demand prediction trigger module S300, permeating the entire remote communication demand prediction process.

[0030] S220. Extract weight parameters from the initial prediction model, perform cross-validation tuning, and generate the optimized model. The system extracts weight parameters from the initial prediction model. Specifically, the initial prediction model is used as the input source, and this model includes a trained LSTM network structure and its corresponding weight parameters. The system first parses and maps the weight parameters to ensure compatibility with subsequent cross-validation tuning. Further, based on a preset cross-validation strategy, the system divides the training dataset into multiple mutually exclusive subsets. One subset is used sequentially as the validation set, and the rest as the training set. The LSTM model is repeatedly trained to evaluate its generalization performance under different data partitions. Specifically, the tuning process includes adjusting hyperparameters such as the number of network layers, the number of hidden units, the learning rate, and the regularization coefficient. The parameter space is systematically explored using methods such as grid search or Bayesian optimization; specifically, Bayesian optimization is used for cross-validation tuning. After each training round, the system calculates the model's performance metrics on the validation set, such as Mean Squared Error (MSE), Mean Absolute Error (MAE), and prediction accuracy, and records the tuning log for performance analysis. For parameter combinations with poor performance, the system automatically adjusts the configuration and repeats training until a preset performance threshold or the maximum number of iterations is reached. After completing all cross-validation rounds, the system synthesizes the results of each round and selects the best-performing weight parameter combination as the final model parameters. Subsequently, the system generates an optimized model based on the optimized weight parameters. Specifically, the weight parameters are integrated with the network structure to form a unified model structure. This optimized model is used as the output field name "Optimized Model" in this step and is passed to the "Deployment Model" in the subsequent step S230 for encapsulation processing. The optimized model not only improves prediction accuracy and stability but also provides a reliable model foundation for subsequent real-time connection demand prediction, which is consistent throughout the machine learning model training and prediction process of this invention.

[0031] In another embodiment, the system acquires the time-series sample structure generated in step S130, which includes multi-dimensional normalized time-series features segmented by a sliding window and their historical connectivity requirement labels. First, data format verification and integrity checks are performed on the time-series sample structure to remove missing or abnormal samples, ensuring the validity of the input data. Then, the input sequence for a Long Short-Term Memory (LSTM) network model is constructed, employing a multi-layer LSTM structure to capture time dependencies. Network parameters are dynamically set according to the feature dimensions, with the input layer dimension aligned to the feature vector length, and the number of hidden layer units determined by an empirical formula. Confirmed, among which Let be the input feature dimension. Formula ① defines the improved cross-entropy loss function:

[0032] in, Improved cross-entropy loss function for LSTM training: Number of samples in the batch (Source: Total number of samples in the time-series sample structure); Sample index (from 1 to ); For the first The true labels of each sample (source: historical connection demand labels, binary encoding); For LSTM Hidden state at time step (Source: Model forward propagation output, dimension) ); For the output layer weights and biases (updated during training after initialization); This is the regularization coefficient (source: hyperparameter configuration table, dimension is proportionality coefficient). Use the Sigmoid activation function; L2 norm squared; Number of hidden layer units in LSTM; : Input feature dimension (from the temporal sample structure).

[0033] Formula ① The optimization objective of backpropagation drives the adjustment of network weights. After training, the initial prediction model, including the network structure and weight parameters, is output and consumed by the "model to be optimized" in S220.

[0034] Furthermore, weight parameters are extracted from the initial prediction model. With the hidden state matrix The training and validation sets are divided based on a cross-validation strategy. Bayesian optimization is used for cross-validation tuning, specifically by dynamically adjusting hyperparameters using a Bayesian optimization method. Formula ② is defined as the weighted performance index for multi-fold cross-validation. in, The weighted performance metric for multi-fold cross-validation (fusing MSE and MAE); For fold index (from 1 to ): The number of folds for cross-validation (source: strategy table configuration, default value 5); For the first The weighting coefficients of the fold (calculated from the proportion of the validation set, satisfying...) ); Mean squared error (source: calculation of validation set prediction results and true labels); The weighting coefficients for MAE (Source: Hyperparameter tuning); For the first Mean absolute error of the fold (calculated by comparing the predicted and actual values ​​on the validation set).

[0035] Formula ② Used to evaluate the generalization performance of different hyperparameter combinations, driving the grid search process. The optimized weight parameters are used to generate the final parameters through the weight fusion strategy in Equation ③:

[0036] in, These are the final weight parameters after fusion; Model index (from 1 to ); The number of optimal parameter combinations; For the first The weights of each model (by...) Inverse proportional distribution, satisfying ); For the first Weight matrices of each model (Source: Candidate Model Pool).

[0037] Formula ③ As a core parameter of the optimized model, it is consumed by the "deployment model" of S230.

[0038] S230. Embedded encapsulation is performed on the optimized model to generate a deployable model structure; The optimized model is embedded and encapsulated. Specifically, the optimized model is used as the input source, and it includes a cross-validated LSTM network structure and weight parameters. The system first converts the model structure and parameters to suit the computational resources and storage limitations of the embedded deployment environment. Specifically, model pruning and weight quantization techniques are used to reduce model size and computational complexity. Further, the system encapsulates the optimized model into an executable deployment unit, including model files, inference engine interfaces, and necessary runtime configuration files, supporting efficient invocation by embedded devices. During the encapsulation process, the system performs consistency verification, including model inference accuracy detection and resource consumption assessment, to ensure stable model operation in the target environment. Subsequently, the system stores the encapsulated model structure as a deployable model structure, which is used as the output field name "Deployable Model Structure" in this step, and passes it to the "Prediction Model" in step S310 for invocation by the connection demand prediction trigger module. The deployable model structure not only achieves efficient model embedding and invocation but also supports rapid prediction response to subsequent real-time data input, which is integral to the remote communication connection demand prediction and dynamic adjustment process of this invention.

[0039] In summary, the technical effects of this step are as follows: Through efficient training of the time-series sample structure, systematic tuning of weight parameters, and embedded encapsulation of the model, a high-performance prediction model is generated, laying the foundation for predicting remote communication connection needs and supporting real-time triggering and dynamic optimization of subsequent modules.

[0040] Step S300 includes at least steps S310-S330: S310. Obtain the real-time communication request, perform feature vector transformation processing, and obtain the input features; The steps of acquiring real-time communication requests, performing feature vector transformation processing, and obtaining input features specifically include the following implementation process. The input features, as the input source, originate from the deployable model structure and real-time communication request data output in step S230. Specifically, the system first receives real-time communication requests initiated by remote communication terminals or the application layer. These requests contain multi-dimensional information such as timestamps, communication types, request frequencies, data packet sizes, terminal identifiers, and network status. Further, the real-time communication request data undergoes preprocessing, including data format validation, missing field completion, and outlier filtering. During preprocessing, the system marks abnormal requests based on a predefined rule set. Anomalies include abnormal request timestamps, excessive data packet sizes, and invalid terminal identifiers. These anomaly records are written to a log file for subsequent analysis. Subsequently, the system maps valid communication request data into feature vectors. Specifically, a feature encoding method is used to convert categorical data into numerical vectors, while numerical data is standardized according to the model input requirements. During feature vector transformation, the system dynamically adjusts feature weights based on historical communication patterns to enhance the expressive power of key features. Furthermore, the system performs weighted fusion of feature vectors based on real-time network environment parameters, such as current signal strength and link quality indicators, to form comprehensive input features. These input features are output in a structured format, containing multi-dimensional numerical vectors and their corresponding time labels. This output field name, "Input Feature," is used in this step and passed to the "Prediction Input" in step S320 for subsequent demand probability prediction. These input features not only provide basic data for real-time connection demand prediction but also permeate the connection demand prediction triggering module of this invention, supporting dynamic adjustment of connection establishment timing.

[0041] S320. Extract time dimension information from input features, perform demand probability prediction, and generate prediction results; The steps of extracting time dimension information from the input features, performing demand probability prediction, and generating prediction results specifically include the following implementation details. The prediction input, as the input source, comes from the input features output in step S310. The system first analyzes the input features in terms of time dimension, specifically identifying the timestamp information and time series features contained in the feature vector, such as request time intervals, peak period markers, and periodic change indicators. Further, the system constructs a time series input based on the time dimension information and inputs it into an embedded Long Short-Term Memory (LSTM) network model. The model structure and weights are derived from the deployable model structure output in step S230. Specifically, the LSTM model captures long-term dependencies and short-term dynamic changes in the time series through its gating mechanism, progressively calculating the hidden states and cell states of the input sequence to achieve probabilistic prediction of future connection demands. During the prediction process, the system adopts a batch inference method, dynamically adjusting the prediction window size and step size in conjunction with the time characteristics of real-time communication requests to adapt to the temporal characteristics of different communication scenarios. Furthermore, the system performs post-processing on the output of the LSTM model, including probability normalization, outlier filtering, and confidence assessment, to ensure the stability and reliability of the prediction results. The prediction results are expressed as demand probabilities, covering the probability distribution of connection demand in the short term, and are used as the output field name "Prediction Result" in this step, and passed to the "Threshold Judgment" module in step S330. The prediction results not only serve as a quantitative indicator of connection demand, but also run through the connection demand prediction process of this invention, supporting dynamic triggering of connection establishment operations.

[0042] S330. Calculate the moving average of the prediction results and generate a trigger command; The step of generating a trigger command by performing a moving average calculation on the prediction results specifically includes the following implementation process. Using the threshold judgment as input, based on the prediction results output in step S320, the system sets the moving average window size and weight coefficients to perform the moving average calculation. Specifically, the system first sets the moving average window size and weight coefficients. The window size is dynamically adjusted based on the time sensitivity of communication needs and the system's response capability, while the weight coefficients use an exponential weighting method to enhance the influence of recent prediction results. Specifically, the system performs a moving average calculation on a continuous sequence of prediction results, calculating the weighted average of the prediction probabilities within the current window point by point, smoothing prediction fluctuations and reducing the impact of occasional anomalies on the triggering decision. Further, the system compares the probability value after moving average with a preset trigger threshold, which is dynamically adjusted based on historical communication load and system resource status. Under conditions that meet or exceed the threshold, the system automatically generates a connection trigger command. The command content includes the trigger time, expected connection type, priority, and related parameters. During the trigger command generation process, the system records the timestamp of the trigger event, the trigger probability value, and related environmental parameters, forming a trigger log for subsequent performance evaluation. Furthermore, the system executes a trigger command cancellation or delay strategy for prediction results that are continuously below a threshold, avoiding frequent triggering that would lead to resource waste. The trigger command is output in a standardized format, including a trigger identifier and control parameters, and is used as the output field name "Trigger Command" in this step, and is passed to the "Connection Command" in step S410 for subsequent communication connection establishment verification. The trigger command is integrated throughout the dynamic connection establishment process of this invention, supporting real-time adjustment of connection initiation timing and resource allocation.

[0043] In summary, the technical effects of this step are as follows: By transforming the feature vector of real-time communication requests, predicting the probability of demand based on the time dimension, and generating moving average trigger instructions, dynamic and accurate prediction and trigger control of connection demand are achieved, laying the foundation for optimizing the timing of remote communication connection establishment.

[0044] Step S400 includes at least steps S410-S430: S410: Obtain the trigger command, initiate the handshake protocol process, and obtain the connection request; The steps of obtaining the trigger instruction, initiating the handshake protocol, and receiving the connection request specifically include the following implementation process. The trigger instruction, as the input source, comes from the "connection instruction" output in step S330. This instruction contains a connection establishment trigger identifier and related parameter information. Specifically, the system first receives the trigger instruction and parses key information such as the connection type, priority, expected connection time, and network environment parameters from the instruction. Further, based on the connection type, the system calls a predefined handshake protocol template, specifically including sub-processes such as connection initialization, authentication, key negotiation, and parameter negotiation. During the handshake protocol initiation process, the system sequentially executes the encapsulation and transmission of the connection request according to the hierarchical structure of the communication protocol stack. Specifically, the handshake request message is encapsulated into a data packet, with necessary header information and a checksum attached, and sent to the target communication node through the underlying transmission interface. After the handshake request is sent, the system starts a timer mechanism to monitor the handshake response timeout and retransmission count. For network anomalies or packet loss, a retry strategy is executed to ensure the integrity of the handshake process. Furthermore, the system monitors the status of each stage of the handshake process in real time, recording the handshake success rate, response latency, and error codes to form a handshake log for subsequent performance analysis and troubleshooting. After the handshake protocol is initiated, the system parses the connection confirmation information, including the connection identifier, session key, and bandwidth allocation parameters, based on the received handshake response message, and generates a connection request object. This connection request is used as the output field name "Connection Request" in this step and is passed to the "Verification Object" in step S420 for subsequent connection quality verification. The connection request not only carries the execution result of the handshake protocol but also provides basic data support for the subsequent communication connection establishment verification module, and is present throughout the dynamic connection establishment process of this invention.

[0045] S420: Extract quality parameters from the connection request, perform bandwidth verification, and generate an available connection; The steps of extracting quality parameters from the connection request, performing bandwidth verification, and generating an available connection specifically include the following implementation details. The connection request, as the input source, comes from the "verification object" output in step S410, and includes a connection identifier, session key, and relevant network quality parameters. The system first parses the quality parameters in the connection request, specifically including key performance indicators such as bandwidth, latency, packet loss rate, and signal-to-noise ratio (SNR). Further, the system verifies the bandwidth indicators based on the current network resource status and Quality of Service (QoS) policy. During the verification process, the system uses a dynamic bandwidth evaluation algorithm, combined with real-time link load, historical bandwidth utilization, and prediction model output, to calculate the actual available bandwidth range. Specifically, the system performs bandwidth allocation verification to determine whether the bandwidth requirement of the connection request meets the current network capacity and priority policy. If not, it triggers a bandwidth adjustment or connection rejection process. The system also performs threshold detection on latency and packet loss rate parameters to ensure that the connection quality meets the preset Service Level Agreement (SLA) requirements. For abnormal or low-quality connection requests, the system performs anomaly marking and logging, including the degree of parameter deviation, the scope of impact, and the processing results, for subsequent network optimization analysis. After verification, the system generates available connection objects. Specifically, connection requests that pass bandwidth verification and meet quality requirements are transformed into available connections, assigned a unique connection identifier and allocated network resource information. The available connection is used as the output field name "Available Connection" in this step and is passed to the "Transmission Channel" in step S430 for subsequent security encryption processing. The available connection not only reflects the quality verification result of the connection request but also provides operable connection resources for the communication connection establishment verification module, and is integral to the connection quality assurance process of this invention.

[0046] S430. Securely encrypt available connections and generate verified connections; The steps for securely encrypting the available connection and generating a verified connection specifically include the following implementation process. The available connection serves as the input source, derived from the "transmission channel" output in step S420, and includes a connection identifier, session key, and allocated network resource information. First, based on the connection identifier, the system calls the corresponding security policy configuration file, specifically including parameters such as encryption algorithm type, key length, authentication mechanism, and encryption mode. Further, the system performs key management operations on the session key, specifically updating, exchanging, and storing the key to ensure its validity and security. During the secure encryption process, the system employs a hybrid encryption scheme combining symmetric and asymmetric encryption. This includes using AES (Advanced Encryption Standard) and ECDSA (Elliptic Curve Digital Signature Algorithm) for secure encryption. Specifically, transmitted data is encrypted using the Advanced Encryption Standard (AES) algorithm, while the key exchange process is protected using the Elliptic Curve Digital Signature Algorithm (ECDSA). During encryption, the system performs data packet encryption and integrity verification. Specifically, a Message Authentication Code (MAC) is appended to the encrypted data packet to prevent data tampering and replay attacks. Furthermore, the system monitors encryption performance metrics such as encryption latency, computational resource consumption, and error rate during encryption, recording encryption logs for security auditing and performance tuning. After encryption is complete, the system generates a verified connection object, containing the encrypted connection identifier, encryption key information, and security authentication status. The verified connection is output as the "Verified Connection" field in this step and passed to the "Transmission Entry" in step S510 for subsequent network topology analysis and path optimization. The verified connection not only ensures the security and integrity of the connection but also provides a secure and reliable connection foundation for the dynamic path optimization transmission module, and is integral to the communication connection establishment and data transmission guarantee process of this invention.

[0047] Steps S410 to S430 initiate the handshake protocol for the connection trigger command, verify the quality parameters of the connection request, and perform secure encryption processing for the available connection. The system completes the core functions of the communication connection establishment and verification module, forming a structured and secure connection entity, which provides a solid foundation for subsequent dynamic path optimization transmission and real-time feedback model updates.

[0048] Step 500 includes at least steps S510-S530: S510. After obtaining the verified connection, perform network topology analysis to obtain the path weight; The steps of obtaining a verified connection, performing network topology analysis, and obtaining path weights specifically include the following implementation process. The verified connection serves as the input source, originating from the "transmission entry" output in step S430, and includes an encrypted connection identifier, encryption key information, and security authentication status. Specifically, the system first receives the verified connection and parses its connection identifier and related network resource information as the foundational data for subsequent network topology analysis. Further, the system accesses a pre-built network topology database containing information such as network nodes, link status, physical location, and link capacity. Specifically, it maps the verified connection's connection identifier to the corresponding network path node. Network topology analysis is performed based on real-time network status. Specifically, the system dynamically collects link utilization, latency metrics, and node load based on the current real-time network status to form a real-time network topology view. The network topology analysis process specifically includes integrating the status of network nodes and links, determining path reachability, and evaluating link quality. The system traverses the network topology to identify possible transmission paths involved in the verified connection. Combining link status information, it calculates the comprehensive weight of each path. The weight indicators cover multiple performance parameters such as link bandwidth, latency, packet loss rate, and node processing capacity. Specifically, the weight calculation uses a weighted summation model, normalizing each performance indicator according to preset weight coefficients before comprehensive calculation to ensure that the weight value reflects the overall quality of the path. Furthermore, the system filters abnormal link states, including link failures, overload, and security risk markers, excluding paths that do not meet transmission requirements. The system also records a network topology analysis log, including the path filtering process, weight calculation results, and abnormal link information, for subsequent optimization and auditing. After completing the network topology analysis, a path weight data structure is generated. Specifically, this structure includes a path identifier, corresponding weight value, and related performance indicators. The path weight, as the output field name of this step, is passed to the "calculation input" in step S520 for dynamic planning calculation. The path weight not only reflects the comprehensive performance evaluation of the current network state but also provides key input for the dynamic path optimization transmission module, running throughout the network adaptive path selection process of this invention.

[0049] S520. Extract link status from path weights, perform dynamic programming calculations, and generate candidate paths; The steps of extracting link states from the path weights, performing dynamic programming calculations, and generating candidate paths specifically include the following implementation details. The path weights, as input sources, come from the "calculation input" output in step S510, and include network path identifiers and their corresponding comprehensive weight values. The system first parses and validates the path weight data, eliminating path records with abnormal weights or missing data to ensure the integrity of the input data. Further, the system performs multi-stage path optimization calculations based on a dynamic programming algorithm, combined with the network topology and path weights. Specifically, the dynamic programming calculation process includes state definition, state transition, and optimal strategy selection. The state is defined as the transmission state of each node in the network, and state transitions are based on link connection relationships and weight values. The system recursively calculates the shortest path or optimal path set from the source node to the target node. The dynamic programming algorithm maintains the cumulative path weights during the calculation process and updates the optimal path strategy in real time. To adapt to dynamic changes in network states, the system introduces a time decay mechanism for path weights, assigning different weights to historical weight data to enhance the timeliness of path selection. Furthermore, the system performs constraint judgments on conflicts and resource competition in path calculation, including link capacity limits, node processing capabilities, and service quality requirements, to ensure that the generated candidate paths meet network resource allocation strategies. During the calculation process, the system records path generation logs, including path selection processes, state transition details, and conflict resolution results, for subsequent performance analysis and troubleshooting. After completing the dynamic programming calculation, the system generates a candidate path list. Specifically, this list contains several path schemes that meet the constraints, and each path records the path node sequence, cumulative weight, and relevant performance parameters. The candidate paths are used as output field names in this step and passed to the "optimization object" in step S530 for delay evaluation. These candidate paths not only provide multiple options for path optimization but also support comprehensive evaluation and final configuration of path performance in subsequent steps, permeating the dynamic path optimization transmission module of this invention.

[0050] S530. Perform delay evaluation on candidate paths to generate the optimal path configuration; The steps for performing delay evaluation on the candidate paths to generate the optimal path configuration specifically include the following implementation process. The candidate paths, as input, are derived from the "optimization object" output in step S520, containing multiple path schemes and their cumulative weights. The system first extracts the path node sequence and corresponding link performance indicators from the candidate paths, specifically involving key parameters such as latency, bandwidth utilization, packet loss rate, and network congestion status. Further, the system performs end-to-end delay evaluation for each candidate path, performing segmented delay accumulation for delay evaluation processing. Specifically, a segmented delay accumulation method is used to summarize the transmission latency, queuing latency, and processing latency of each link in the path to form a total path delay indicator. During the delay evaluation process, the system dynamically adjusts the link latency parameters based on real-time network monitoring data to reflect current network load and status changes. The system also considers the path stability indicators, evaluating the availability and volatility of the path within the expected transmission time, and calculating the path stability score by statistically analyzing historical path performance fluctuations. For multiple candidate paths, the system constructs a comprehensive evaluation model, integrating the delay indicator and stability score, and uses a multi-objective optimization method to rank the paths. Specifically, the system normalizes each indicator according to preset weighting coefficients, calculates the comprehensive path score, and prioritizes the path with the highest score as the optimal path. Further, the system performs path configuration generation, specifically encapsulating the node sequence and related parameters of the selected optimal path into a path configuration file, including path identifiers, link order, bandwidth allocation, and security policies. After path configuration generation, the system performs consistency verification on the configuration file, verifying the path connectivity and the rationality of resource allocation, and eliminating potential configuration conflicts. The system records latency assessment and path configuration logs, including assessment process data, path selection basis, and configuration generation details, for subsequent performance monitoring and optimization. After completing path configuration generation, the system outputs the optimal path configuration as the output field name for this step, which is passed to the "Transmission Log" in step S610 for subsequent transmission process monitoring. The optimal path configuration not only reflects the low latency and high stability characteristics of the path but also provides an efficient network path selection scheme for real-time data transmission, permeating the dynamic path optimization transmission and real-time feedback model update process of this invention.

[0051] In summary, the technical effects of this step are as follows: By analyzing the network topology of the verified connection, calculating the path based on dynamic programming, and evaluating the latency of multiple indicators, the optimal path generation with adaptive network state is achieved, thereby improving the transmission efficiency and stability of remote communication.

[0052] In another embodiment, in S510, the network topology data of the verified connection is obtained, and the path nodes and link states are parsed. A weighted summation model is used to calculate the path weight, and formula ④ is defined as the path score fused with multiple indicators: in, Path weight scoring (result of multi-indicator fusion); Link bandwidth (Source: Network Topology Database); For latency (Source: Real-time monitoring data); Packet loss rate (Source: QoS policy table); Weighting coefficients (Source: Strategy Configuration, satisfying...) ); This is the normalized baseline value (source: historical peak data).

[0053] Furthermore, in S520, link states are extracted from path weights, dynamic programming calculations are performed, and candidate paths are generated, as shown in formula ④. The path weights are used as the basis for calculation and input into the dynamic programming algorithm. The improved Bellman equation is used for dynamic programming calculation; specifically, the optimal path is solved based on the improved Bellman equation (Equation ⑤).

[0054] in, For nodes The minimum cumulative cost; For the current node (source: topology node list); Adjacent nodes (belonging to) ); for The set of adjacent nodes; The edge cost (from formula ④) (Inverse proportional mapping).

[0055] Furthermore, in S530, a delayed evaluation process is applied to the candidate paths to generate the optimal path configuration. Equation ⑤ generates the candidate path set through recursive calculation. Finally, the optimal path is selected through multi-objective ranking using Equation ⑥. in, Identify the optimal path; Candidate paths (belonging to) ); For the candidate path set; For path Weighted scoring (Source: Formula ④); Path stability score (Source: Historical volatility calculation); Weighting factors (source: strategy table configuration); The parameter corresponding to the maximum value (here, it means selecting the path with the highest score).

[0056] Specifically, formula ⑥ The optimal path configuration is generated and consumed by the S610's "transmission log".

[0057] This section summarizes the technical effects: A high-precision prediction model was constructed through LSTM training of time-series samples, weight optimization, and model encapsulation; based on dynamic planning and latency evaluation of network topology, a closed-loop path optimization was achieved, improving communication transmission efficiency and stability.

[0058] Step S600 includes at least steps S610-S630: S610: Obtain the optimal path configuration, perform transmission process monitoring and processing, and obtain performance logs; The steps of obtaining the optimal path configuration, monitoring the transmission process, and generating performance logs specifically include the following implementation process. The optimal path configuration serves as the input source, derived from the "transmission log" output in step S530. This configuration includes path identifiers, path node sequences, bandwidth allocation parameters, and security policy information. Specifically, the system first initializes the transmission module, establishes a data transmission channel, and configures path-related network device parameters based on the optimal path configuration, including routing table updates, link bandwidth allocation, and security authentication settings. During transmission, the system monitors and collects performance indicators. Specifically, the system employs a multi-layered monitoring mechanism to collect real-time data packet transmission, reception, and forwarding status, covering key performance indicators such as transmission latency, packet loss rate, throughput, signal strength, and link stability. Specifically, the system continuously acquires the status information of the transmission channel through network probes and a monitoring module embedded in the protocol stack, forming a multi-dimensional performance data stream. Furthermore, the system performs time-series alignment and anomaly detection on the collected performance data, using a sliding window statistical method to calculate the mean, variance, and peak value of each indicator, identifying abnormal fluctuations and sudden events during the transmission process. Abnormal events include, but are not limited to, link interruption, abnormal increase in packet retransmission count, and security authentication failure. The system records these abnormal events and their timestamps in detail in the performance log. The system also performs real-time compression and storage management of performance data to ensure the integrity and efficient querying of the performance log. The performance log is stored in a structured format, containing timestamps, performance metric values, and descriptions of abnormal events. This is used as the output field name "Performance Log" in this step and passed to the "Analysis Data" in step S620 for subsequent error metric extraction and model updates. The performance log not only provides transmission quality feedback for the dynamic path optimization transmission module but also runs throughout the real-time feedback model update process of this invention.

[0059] S620: Extract error metrics from performance logs, perform incremental learning processing, and generate model updates; The steps of extracting error metrics from the performance logs, performing incremental learning, and generating model updates specifically include the following implementation details. The performance logs, as the input source, are derived from the "analysis data" output in step S610 and contain multi-dimensional performance metrics and abnormal event records during transmission. The system first cleans and preprocesses the performance logs, removing invalid data and abnormal noise, and uses a time synchronization mechanism to align logs across devices, forming a unified performance dataset. Further, based on a predefined error metric system, the system extracts key error metrics from the performance dataset, specifically including transmission delay deviation, packet loss rate fluctuation, throughput decrease, and security event frequency. Error metric calculation employs statistical analysis and signal processing techniques, combining a sliding window and Exponentially Weighted Moving Average (EWMA) method to capture performance trends and sudden anomalies. The extracted error metrics are normalized and mapped to the input space required for model updates. The system then inputs the error metric into the incremental learning module, which, based on the previously trained Long Short-Term Memory (LSTM) network model structure, uses an online learning algorithm to locally adjust the model parameters. Specifically, the incremental learning process includes error backpropagation, gradient update, and weight fine-tuning, combined with an adaptive learning rate strategy to dynamically adjust the model's response speed to new data. The system introduces a regularization term and a catastrophic forgetting mitigation mechanism during the incremental learning process to maintain the model's stability and generalization ability. The incremental learning module records training logs during processing, including changes in the error metric, parameter update magnitudes, and model performance evaluation results, for subsequent performance monitoring and optimization. After completing incremental learning, the system outputs a model update object, specifically containing the fine-tuned network weights and corresponding model structure adjustment information, as the output field name "Model Update" for this step, and passes it to the "Parameter Input" in step S630 for online deployment processing. This model update not only reflects the real-time impact of transmission performance feedback on the model but also runs throughout the machine learning model training and real-time feedback update process of this invention.

[0060] S630. Perform online deployment processing for model updates and generate the updated model; The online deployment process for the model update, specifically generating the updated model, includes the following steps: The model update, as the input source, comes from the "parameter input" output in step S620, containing network weights and structural adjustment information adjusted through incremental learning. The system first performs integrity verification and compatibility testing on the model update, specifically verifying the weight parameter format, model structure consistency, and version matching to ensure seamless integration of the model update into the existing deployment environment. Further, based on embedded deployment requirements, the system performs necessary format conversions and optimizations on the model update, including weight quantization, model pruning, and memory layout adjustments, reducing computational resource consumption and storage usage during model runtime. During online deployment, the system employs a seamless switching strategy and a hot update mechanism for online deployment. Specifically, through parallel dual-model operation and a hot update mechanism, smooth model switching is achieved, avoiding service interruptions. The system performs model performance verification before and after deployment, using a preset test set and real-time performance monitoring data to evaluate the prediction accuracy, response speed, and resource consumption of the updated model, ensuring that the model update meets system operating requirements. After deployment, the system stores the updated model in the model management repository, including version information, update logs, and performance metrics, supporting subsequent version rollback and audit trails. The updated model is used as the output field name "Updated Model" in this step and is passed to the "Training Data" in step S210 for continuous iteration of subsequent machine learning model training. The updated model is integrated throughout the machine learning model training module and the real-time feedback model update module of this invention, forming a closed-loop model optimization and application mechanism.

[0061] In summary, the technical effects of this step are as follows: Through in-depth analysis of transmission performance logs, incremental learning based on error indicators, and online deployment processing, the model is dynamically updated in real time, improving the prediction accuracy and adaptive capability of the remote communication system.

[0062] Example 2: Figure 2 A structural block diagram of a predictive remote communication system according to an embodiment of the present invention is shown. Figure 2 As shown, the structure may include: Node configuration module 01 is used to acquire communication node configuration parameters and complete network topology constraints. Specifically, it receives the input communication node configuration parameters from the system initialization phase, and under the configured network topology constraints, completes the setting of node connection structures, calibration of interface parameters, and initialization of links to form a node topology diagram. The node topology diagram is recorded as a topology structure field and maintains a consistent association with the current network state. The topology structure field is transmitted to the data acquisition module for use as network structure information, while the topology log is retained for subsequent network state monitoring.

[0063] Data acquisition module 02 is used to collect real-time transmission data packets and historical delay records. Specifically, it receives network structure information output from the node configuration module and real-time transmission data packets on the communication link, performs data packet parsing, traffic statistics, and anomaly filtering on the real-time transmission data packets, and forms historical delay records based on the historical record policy table. The historical delay records are passed as delay data fields to the feature processing module for invocation, and the corresponding timestamp information is registered in the data buffer for the feature processing module to read.

[0064] The feature processing module 03 is used to perform feature extraction processing on the real-time transmitted data packets and normalize the historical delay records. Specifically, it receives delay data fields from the data acquisition module, performs feature extraction and normalization processing in conjunction with the real-time transmitted data packets, and generates a transmission feature vector when the feature extraction rules are met. The transmission feature vector is then passed to the synchronization module as a feature input field, and the correspondence between the feature vector and the feature extraction parameters is recorded in the feature storage area.

[0065] Synchronization module 04 is used to perform timestamp alignment processing on the transmission feature vector and the standard delay sequence. Specifically, it performs timestamp parsing and alignment processing based on the feature input fields from the feature processing module to generate an aligned feature sequence; the aligned feature sequence is output to the arbitration module, and the alignment status is returned to the feature processing module for registration.

[0066] Arbitration module 05 is used to perform link congestion arbitration based on a dynamic threshold policy table. Specifically, it receives the alignment feature sequence from the synchronization module and the dynamic threshold policy table from the policy management unit, performs congestion determination, priority calculation, and path selection to obtain the arbitration result; the arbitration result is provided to the control module as a path decision field, and an index relationship consistent with the time series is maintained.

[0067] Control module 06 is used to output path switching instructions to the routing controller. Specifically, it receives the path decision field from the arbitration module, performs instruction generation and signal encoding on the path decision field to form a path switching instruction; outputs the path switching instruction to the routing controller interface, and sends back an effectiveness confirmation to the update module for registration.

[0068] The update module 07 is used to record the effective timestamp and update the dynamic threshold policy table. Specifically, it receives the effective confirmation from the control module and the policy log from the policy management unit, performs timestamp recording, policy update and version control to obtain the updated policy table; the updated policy table is returned to the node configuration module for parameter backfeeding, and the policy records are updated.

Claims

1. A method for predicting remote communication, characterized in that, include: Historical communication data is acquired, outliers are identified and cleaned based on box plots or standard deviations, normalization is performed using min-max normalization or Z-score normalization, and sliding window segmentation is performed by setting the sliding window size and step size according to the time granularity of communication requirements to generate time series sample structures. Obtain the temporal sample structure, train a multi-layer LSTM model, determine the number of hidden layer units through empirical formulas, extract weight parameters, perform cross-validation tuning using Bayesian optimization, and perform model pruning and weight quantization to encapsulate a deployable model structure. It receives real-time communication requests, performs feature encoding and standardization when transforming feature vectors, uses an LSTM model to perform demand probability prediction, sets the sliding window size and weight coefficients to perform moving average calculation, and generates trigger commands. Initiate a handshake protocol that includes connection initialization and authentication based on the trigger command, perform bandwidth verification based on real-time link load, and use AES and ECDSA for security encryption to generate a verified connection; After obtaining the verified connection, network topology analysis is performed based on the real-time network status. The improved Bellman equation is used for dynamic programming calculation, and segmented delay accumulation is performed for delay evaluation to generate the optimal path configuration. The system monitors the transmission process to collect performance metrics, extracts error metrics including transmission delay deviation and packet loss rate fluctuations, performs incremental learning, and uses a hot update mechanism for online deployment to generate an updated model.

2. The predictive remote communication method according to claim 1, characterized in that, The process of obtaining the temporal sample structure and training a multi-layer LSTM model specifically includes: Perform data format verification and integrity checks on the time series sample structure to exclude missing or abnormal samples; The input sequence for constructing the Long Short-Term Memory Network Model is adopted, and a multi-layer LSTM structure is used to capture long-term dependencies and nonlinear dynamic changes in the time series. Batch training is performed based on time-series sample structure, and the network weight parameters are iteratively adjusted by combining backpropagation algorithm with gradient descent optimization method. Early stopping mechanism and regularization technique are introduced to prevent overfitting. At the same time, cross-validation is used to divide the training set and validation set, and the model performance indicators such as loss function value and prediction error are dynamically monitored.

3. The predictive remote communication method according to claim 1, characterized in that, The process of extracting weight parameters and performing cross-validation tuning using Bayesian optimization specifically includes: Based on the preset cross-validation strategy, the training dataset is divided into multiple mutually exclusive subsets. One subset is used as the validation set and the rest are used as the training set. The LSTM model is trained repeatedly. The tuning process includes adjusting the number of network layers, the number of hidden units, the learning rate, and the regularization coefficient hyperparameters, and systematically exploring the parameter space through Bayesian optimization methods; After each round of training, calculate the model's performance metrics on the validation set and record tuning logs for performance analysis. After completing all rounds of cross-validation, the best-performing combination of weight parameters is selected as the final model parameters by combining the results of each round.

4. The predictive remote communication method according to claim 1, characterized in that, The process of performing model pruning and weight quantization and encapsulating it into a deployable model structure includes: The model structure and parameters are converted to adapt to the computing resources and storage limitations of the embedded deployment environment. Model pruning and weight quantization techniques are used to reduce model size and computational complexity. The optimized model is packaged into an executable deployment unit, which includes model files, inference engine interfaces, and necessary runtime configuration files; During the encapsulation process, consistency verification is performed, including model inference accuracy detection and resource consumption assessment.

5. The predictive remote communication method according to claim 1, characterized in that, The process of receiving real-time communication requests and performing feature encoding and standardization when transforming feature vectors includes: Receive real-time communication requests initiated by remote communication terminals or application layers. These requests include timestamps, communication types, request frequencies, data packet sizes, terminal identifiers, and network status information. Preprocess the real-time communication request data, including data format validation, missing field completion, and outlier filtering; Valid communication request data is mapped to feature vectors. Feature encoding methods are used to convert categorical data into numerical vectors, and numerical data is standardized according to the model input requirements.

6. The predictive remote communication method according to claim 1, characterized in that, The process of performing demand probability forecasting using an LSTM model includes: The input features are analyzed in terms of time dimension to identify timestamp information and time series features contained in the feature vector; A time series input is constructed based on time dimension information and fed into an embedded long short-term memory network model; LSTM models capture long-term dependencies and short-term dynamic changes in time series through their gating mechanism, and progressively calculate the hidden state and cell state of the input sequence to achieve probabilistic prediction of future connection needs. The output of the LSTM model is post-processed, including probability normalization, outlier filtering, and confidence assessment.

7. The predictive remote communication method according to claim 1, characterized in that, The process of setting the sliding window size and weighting coefficients, performing a moving average calculation, and generating trigger instructions includes: The size and weighting coefficients of the moving average window are set. The window size is dynamically adjusted according to the time sensitivity of communication needs and the system response capability. The weighting coefficients adopt an exponential weighting method to enhance the influence of recent prediction results. Perform a moving average calculation on the continuous sequence of prediction results, and calculate the weighted average of the prediction probabilities within the current window point by point; The probability value after the moving average is compared with the preset trigger threshold, which is dynamically adjusted according to historical communication load and system resource status. Under conditions that are met or exceeded, a connection trigger command is automatically generated. The command includes the trigger time, expected connection type, priority, and related parameters.

8. The predictive remote communication method according to claim 1 or 2, characterized in that, The training of a multi-layer LSTM model involves determining the number of hidden layer units using an empirical formula, extracting weight parameters, and employing Bayesian optimization for cross-validation tuning. The expression for this optimization includes: Network parameters are dynamically set based on the feature dimension, the input layer dimension is aligned with the feature vector length, and the number of hidden layer units is determined by an empirical formula. Confirmed, among which Define an improved cross-entropy loss function for the input feature dimension: ； in, Improved cross-entropy loss function for LSTM training: This represents the batch sample size. Sample index; For the first The true label of each sample; For LSTM The hidden state of the time step; For output layer weights and biases; The regularization coefficient is used. Use the Sigmoid activation function; L2 norm squared; Number of hidden layer units in LSTM; As the optimization objective of backpropagation, it drives the adjustment of network weights and outputs the initial prediction model, which includes the network structure and weight parameters. Furthermore, weight parameters are extracted from the initial prediction model. With the hidden state matrix The training and validation sets are divided based on a cross-validation strategy. Bayesian optimization is used to fine-tune the cross-validation, specifically by dynamically adjusting hyperparameters using Bayesian optimization and defining a weighted performance metric for multi-fold cross-validation. in, The weighted performance metric for multi-fold cross-validation; For fold index: For cross-validation folds; For the first The weighting coefficient of the fold; Mean square error; These are the weighting coefficients for MAE; For the first The mean absolute error of the fold; Used to evaluate the generalization performance of different hyperparameter combinations, driving the grid search process, the optimized weight parameters are used to generate the final parameters through a weight fusion strategy: ； in, These are the final weight parameters after fusion; For model indexing; The number of optimal parameter combinations; For the first The weights of each model; For the first The weight matrix of each model.

9. The predictive remote communication method according to claim 1, characterized in that, The improved Bellman equation is used for dynamic programming calculations, and piecewise delay accumulation is performed for delay evaluation. The expression for the optimal path configuration is as follows: The path weights are calculated using a weighted summation model, and a path score integrating multiple indicators is defined: in, Score the path weight; Link bandwidth; For time delay; Packet loss rate; These are the weighting coefficients; This is the normalized baseline value; Furthermore, The path weights are used as the basis for calculation and input into the dynamic programming algorithm. The improved Bellman equation is used for dynamic programming calculation to solve for the optimal path. ； in, For nodes The minimum cumulative cost; For the current node; Adjacent nodes; for The set of adjacent nodes; For edge cost; Furthermore, the optimal path is selected through multi-objective ranking: in, Identify the optimal path; Candidate paths; For the candidate path set; For path Weighted scoring; Path stability score; As a weighting factor; The parameter corresponding to the maximum value.

10. A predictive remote communication system, applied to the method of any one of claims 1-9, characterized in that, include: The node configuration module is used to obtain communication node configuration parameters and complete network topology constraints. The data acquisition module is used to collect real-time transmitted data packets and historical delay records; The feature processing module is used to extract features from real-time transmitted data packets and normalize historical delay records. The synchronization module is used to timestamp-align the transmitted feature vector with the standard delay sequence; The arbitration module is used to perform link congestion arbitration based on a dynamic threshold policy table; The control module is used to output path switching instructions to the routing controller; The update module is used to record the effective timestamp and update the dynamic threshold strategy table.

Images