Dual-mode communication optimization system and method thereof

Abstract

The invention provides a dual-mode communication optimization system and method, and belongs to the technical field of communication. The method for judging the upgrading mode of the dual-mode communication model considering the long-term performance extreme value is provided, accurate grading judgment of model upgrading requirements is achieved, unnecessary full-amount training overhead is avoided, and stability and robustness of a system in long-term operation are guaranteed. Secondly, after upgrading mode judgment of the dual-mode communication model is completed, in order to further improve the optimization efficiency and pertinence of a system in a complex channel environment, a dual-mode communication model upgrading optimization method based on historical performance gain is provided, an optimization strategy can be adaptively adjusted according to the intensity of environment change, rapid convergence is achieved during performance degradation, and the optimization efficiency and pertinence of the system in a complex channel environment are improved. And accurate focusing is realized during local fine tuning, so that efficient evolution of a dual-mode communication model is realized.

Description

Technical Field

[0001] This invention provides a dual-mode communication optimization system and method, belonging to the field of communication technology. Background Technology

[0002] Dual-mode communication systems, as a key technology for achieving adaptive switching between low latency and high throughput, are crucial for ensuring the stability, reliability, and resource utilization efficiency of communication networks. With the large-scale access of heterogeneous devices from multiple sources, such as vehicle-to-everything (V2X), industrial IoT, and smart power terminals, the dynamism of communication links and the diversity of task types have significantly increased. Real-time fluctuations in channel status, service priorities, and network load place higher demands on the adaptability of communication systems. Existing communication models often employ fixed structures or static parameters for mode selection and channel allocation, making it difficult to dynamically adjust to changes in the channel environment and service priorities. This leads to performance degradation, increased latency fluctuations, and uneven resource allocation during long-term operation. Therefore, there is an urgent need for a method that can adaptively judge and optimize model upgrade strategies based on terminal operating status and channel change trends to achieve long-term performance maintenance and stable operation of dual-mode communication systems in complex environments.

[0003] In existing research on dual-mode communication optimization, model training and upgrades are mostly controlled based on fixed time intervals or static thresholds, lacking dynamic perception of terminal performance evolution trends and long-term performance extrema, making it difficult to accurately determine the timing and scope of model upgrades. Furthermore, existing optimization methods often model mode selection and channel selection independently, failing to effectively integrate their synergistic relationship, resulting in low model training efficiency, slow convergence speed, and susceptibility to overfitting or performance fluctuations when the communication environment changes rapidly. Simultaneously, some algorithms do not fully utilize historical performance gain information during optimization, failing to balance training computation and long-term performance gains, thus limiting the widespread application of models in dynamic channel scenarios. In summary, establishing an adaptive optimization method for dual-mode communication models that integrates performance prediction, long-term extrema calculation, and historical gain analysis has become a key technical issue for improving system communication performance and intelligence. Summary of the Invention

[0004] To address the aforementioned issues, this invention proposes a method for determining the upgrade mode of a dual-mode communication model that considers long-term performance extrema. First, a dual-mode communication model is constructed on the end side, and communication links are jointly decided through a mode selection sub-network and a channel selection sub-network. Second, a sliding time window mechanism is used to collect performance data and perform rolling predictions to obtain performance sequences for future periods. Then, performance extrema are calculated from both temporal and spatial dimensions. The vertical extrema incorporate a second-order difference term to quantify the oscillation severity of a single terminal, while the horizontal extrema integrate equilibrium and constraint violation frequencies to quantify the system's adaptability defects. Finally, a system-level comprehensive index is constructed based on the mean and variance of each terminal's extrema. An upgrade strategy is adaptively selected through a dual-threshold judgment logic: a hard upgrade is performed to reshape the structure when the system oscillates; a soft upgrade is performed to fine-tune the channel when the system is unbalanced; and the status quo is maintained when the system is stable and balanced. This method achieves accurate hierarchical determination of model upgrade needs, avoiding unnecessary full-scale training overhead while ensuring the stability and robustness of the system in long-term operation.

[0005] Secondly, after determining the upgrade mode of the dual-mode communication model, this invention proposes an upgrade optimization method for the dual-mode communication model based on historical performance gains to further improve the optimization efficiency and targeting of the system in complex channel environments. First, a differentiated sample selection mechanism based on a historical experience pool is constructed. According to the current upgrade mode, sampling weights are dynamically calculated using performance gains and scene similarity. High-value samples are selected from historical records and fused with real-time observation samples to form a training set. Second, a parameter update mask mechanism is introduced to construct a weighted loss function. Differential updates are achieved under a unified optimization framework. In hard upgrade mode, all parameters are jointly trained. In soft upgrade mode, modal parameters are forcibly locked through a mask matrix, and only channel parameters are fine-tuned. Finally, the optimized parameters are sent to the end-user for synchronization. This method can adaptively adjust the optimization strategy according to the severity of environmental changes, converge quickly when performance degrades, and focus precisely during local fine-tuning, thereby achieving efficient evolution of the dual-mode communication model.

[0006] To facilitate the practical application of the aforementioned technologies, this invention proposes a dual-mode communication optimization system. This system includes a model synchronization module, a dual-mode communication model decision-making module, a performance data acquisition module, and a model upload module deployed on the end side; and a performance prediction and extreme value analysis module, an upgrade mode judgment module, a model upgrade optimization module based on historical performance gains, and a model distribution module deployed on the edge side. This system effectively integrates the entire process of real-time data acquisition on the end side with long-term performance prediction on the edge side, multi-dimensional extreme value analysis, graded determination of upgrade modes, and mask differentiation optimization. It can intelligently decide on update strategies and parameter ranges based on changes in the channel environment using an end-edge interaction mechanism. This significantly improves the long-term stability and robustness of the dual-mode communication system while greatly reducing computational overhead by avoiding invalid training, achieving an optimal balance between overall communication efficiency and resource utilization.

[0007] The specific technical solution is as follows:

[0008] A dual-mode communication optimization system includes a model synchronization module, a dual-mode communication model decision module, a performance data acquisition module, and a model upload module deployed on the end side, as well as a performance prediction and extreme value analysis module, an upgrade mode judgment module, a model upgrade optimization module based on historical performance gains, and a model distribution module deployed on the edge side. The system comprises several modules: a model synchronization module, working in conjunction with a dual-mode communication model decision module, performs hot loading upon receiving side-side parameters and achieves real-time joint decision-making at the end-side via a mode and channel selection sub-network; a performance data acquisition module and a model upload module, responsible for collecting operational data and establishing uplink channels to upload historical performance sequences and original parameters to the side-side; a performance prediction and extreme value analysis module, based on a rolling prediction mechanism, generates performance sequences for future periods and quantifies the temporal longitudinal oscillation stability of a single terminal and the spatial lateral collaborative adaptability of multiple terminals; an upgrade mode judgment module, based on the statistical characteristics of extreme values, constructs a system-level index to dynamically distinguish between hard update, soft update, and no update modes, enabling accurate determination of model upgrade requirements; a model upgrade optimization module based on historical performance gains utilizes differentiated sample screening and parameter update masking mechanisms to achieve adaptive full reshaping or local fine-tuning of mode and channel parameters; and a model distribution module, after optimization, securely distributes the updated model parameters to the end-side, achieving side-end collaboration and closed-loop adaptive updates.

[0009] Specifically, it includes end-side subsystems and edge-side subsystems;

[0010] The end-side subsystem includes a model synchronization module, a dual-mode communication model decision module, a performance data acquisition module, and a model upload module.

[0011] The side subsystem includes a performance prediction and extreme value analysis module, an upgrade mode judgment module, a model upgrade optimization module based on historical performance gains, and a model distribution module;

[0012] The details are as follows:

[0013] The model synchronization module receives updated dual-mode communication model parameters from the edge and performs parameter replacement and hot loading on the local model. This module ensures that the edge-side decision model is consistent with the edge-side optimization results, and its updated model parameters are input to the dual-mode communication model decision module.

[0014] The dual-mode communication model decision module is used to jointly decide on mode selection and channel selection at the endpoint based on the synchronized dual-mode communication model, given channel state and data priority. This module consists of two parts: a mode selection subnetwork and a channel selection subnetwork. It takes state information as input and outputs the optimal communication mode and channel selection results. Its operational data is input to the performance data acquisition module.

[0015] The performance data acquisition module is used to collect communication performance data from the end-user device in real time during operation and construct a performance time series. This module records raw data such as latency and packet loss rate in each time slot, providing raw data support for subsequent analysis. The acquisition results are input into the model upload module.

[0016] The model upload module is used to establish an uplink data transmission channel from the endpoint to the edge. This module is responsible for packaging and uploading the collected historical performance sequences and the current original dual-mode communication model parameters of the endpoint, and outputting them to the performance prediction and extreme value analysis module and the model upgrade and optimization module based on historical performance gains of the edge subsystem, respectively.

[0017] The performance prediction and extreme value analysis module is used to obtain the performance prediction sequence for future periods by using a rolling prediction mechanism and extracting multi-dimensional features. This module calculates the longitudinal performance extreme value of a single terminal to characterize temporal stability and calculates the lateral performance extreme value of multiple terminals to characterize spatial balance. The output quantitative index is input to the upgrade mode judgment module.

[0018] The upgrade mode determination module is used to determine the system's model upgrade requirements based on statistical methods. This module statistically analyzes the extreme values ​​of each terminal to construct a comprehensive longitudinal instability index and a lateral adaptation index. When the longitudinal instability index exceeds the limit, a structural failure is determined, triggering a hard upgrade mode; when the longitudinal instability index is stable but the lateral adaptation index exceeds the limit, local adaptability is determined, triggering a soft upgrade mode; otherwise, no upgrade is performed. The determination results of this module are input to the model upgrade optimization module based on historical performance gains.

[0019] The model upgrade and optimization module based on historical performance gains receives the original model parameters and performs differentiated model training. This module maintains a historical experience pool and dynamically adjusts sample sampling weights to construct the training set based on the upgrade mode, performance gain, and scene similarity. It also introduces a parameter update mask mechanism: in hard upgrade mode, all parameters are updated; in soft upgrade mode, mode selection parameters are locked through a mask, and only channel selection parameters are optimized. The optimized model parameters are then input to the model distribution module.

[0020] The model distribution module encapsulates the updated model parameters after model optimization and distributes them to the corresponding edge subsystems via a downlink channel. This module transmits parameters through a secure communication channel, ensuring data integrity and model consistency, and enabling data input and closed-loop updates for the edge model synchronization module.

[0021] A dual-mode communication optimization method includes the following steps:

[0022] S1: Determination of upgrade mode for dual-mode communication model considering long-term performance extremes

[0023] First, a dual-mode communication model is constructed at the terminal side, and the communication link is jointly decided by a mode selection subnetwork and a channel selection subnetwork. Second, a sliding time window mechanism is used to collect performance data and perform rolling predictions to obtain the performance sequence for future time periods. Then, performance extrema are calculated from both the time and spatial domains. The vertical extrema introduce a second-order difference term to quantify the oscillation severity of a single terminal, while the horizontal extrema integrate equilibrium and constraint violation frequencies to quantify the system's adaptability defects. Finally, a system-level comprehensive index is constructed based on the mean and variance of the extrema of each terminal, and an upgrade strategy is adaptively selected through a dual-threshold judgment logic: a hard upgrade is performed to reshape the structure when the system oscillates, a soft upgrade is performed to fine-tune the channel when the system is unbalanced, and the status quo is maintained when the system is stable and in equilibrium.

[0024] S2: Upgrade and optimization of the dual-mode communication model based on historical performance gains

[0025] First, a differentiated sample selection mechanism based on a historical experience pool is constructed. According to the current upgrade mode, sampling weights are dynamically calculated using performance gain and scene similarity. High-value samples are selected from historical records and fused with real-time observation samples to form a training set. Second, a parameter update mask mechanism is introduced to construct a weighted loss function. Differentiated updates are achieved within a unified optimization framework. In hard upgrade mode, all parameters are jointly trained. In soft upgrade mode, modal parameters are forcibly locked using a mask matrix, and only channel parameters are fine-tuned. Finally, the optimized parameters are distributed to the edge device to complete synchronization.

[0026] Specifically:

[0027] S1: Determination of upgrade mode for dual-mode communication model considering long-term performance extremes

[0028] S1.1 End-side dual-mode communication model decision

[0029] Considering There are 1 terminal, and the set is denoted as _ . ,terminal Shared with edge servers Channels, represented as a set All channels are divided into two subsets: the set of channels corresponding to the first mode. The channel set corresponding to the second mode ,and .terminal The decision variables include modal selection variables. and channel selection variables ,when Time indicates terminal Selecting the first communication mode, the terminal from... Select Channel ; Time indicates terminal Selecting the second communication mode, the terminal from Select Channel .

[0030] Each terminal is equipped with a dual-mode communication model. ,in For model parameters, and Representing terminals respectively Mode selection subnetwork parameters and channel selection subnetwork parameters. This is status information, where For channel state, Prioritize the data.

[0031] By input The dual-mode communication model makes a decision and outputs the selected communication mode and the corresponding channel, which is represented as follows;

[0032] (1)

[0033] In the formula, The modality selection result is determined by the parameters of the modality selection subnetwork. The channel selection result is determined by the parameters of the channel selection sub-network.

[0034] S1.2 Performance Data Acquisition and Statistics

[0035] No. Time slot, acquisition terminal latency With packet loss rate Construct a performance loss function , is represented as:

[0036] (2)

[0037] In the formula, and Do not represent the weights of latency and packet loss rate.

[0038] Set the observation window length to .exist Time slots, for terminals Construct historical performance sequences , is represented as:

[0039] (3)

[0040] Pre-trained prediction model Or GRU, to perform The prediction process employs an iterative prediction mechanism, which involves a "prediction-filling-re-prediction" rolling process. The specific steps are as follows:

[0041] (1). Single-step prediction: using a pre-trained prediction model Predict the output of the next time slot Performance prediction , is represented as:

[0042] (4)

[0043] (2). Rolling update: The predicted values ​​are updated through a first-in-first-out rolling operation. Replace historical performance sequence The original first data A new historical performance sequence is obtained. And input the prediction model again. Make predictions.

[0044] Repeat the above steps Next, through rolling prediction, a length of [length] is obtained. The prediction performance sequence is represented as:

[0045] (5)

[0046] The sequence It is then output to the performance extreme value comprehensive analysis module for subsequent extreme value feature extraction.

[0047] S1.3 Comprehensive Analysis of Horizontal and Vertical Performance Extremes

[0048] Define longitudinal performance extreme values This reflects the degree of performance fluctuation of a single terminal over a future period. A second-order difference term is introduced to identify non-stationary states that may lead to frequent mode switching. Specifically, it is expressed as:

[0049] (6)

[0050] In the formula, and The weights are the fluctuation amplitude weight and the oscillation acceleration weight, respectively, which can be determined by calculating the information entropy using the entropy weight method based on historical data. For the terminal The predicted average loss is expressed as The first term is the standard deviation, representing the overall fluctuation of the terminal's performance. The second term is the sum of the absolute values ​​of the second-order differences, representing the acceleration of the change in the terminal's performance. For second-order difference operations, it is represented as:

[0051] (7)

[0052] Horizontal performance extreme values ​​are constructed from three dimensions: indicator balance, constraint violation frequency, and constraint violation distribution differences. .

[0053] First, statistical computing terminal In the prediction window Number of times the constraint is not satisfied , is represented as:

[0054] (8)

[0055] In the formula, The performance threshold is determined based on the maximum allowable latency and packet loss rate for the current terminal service type, combined with the aforementioned performance loss function. This is an indicator function; it takes the value 1 if the condition is met, and 0 otherwise.

[0056] Next, calculate Lateral performance extreme values ​​of time slots , is represented as:

[0057] (9)

[0058] In the formula, the first term represents the performance distribution uniformity, reflecting the dispersion of the average performance loss of different terminals during the prediction period. The first term represents the average performance loss across all terminals. The second term represents the system-level constraint violation strength, reflecting the severity of the system's inability to meet performance constraints within a future period. This value normalizes the number of terminals and the time window, directly reflecting the overall unavailability probability of the system. The third term represents the violation frequency variability, reflecting the uneven distribution of violation phenomena across different terminals. The larger this value, the worse the model's adaptability to individual terminals. These are weights for performance distribution balance, system-level constraint violation intensity, and violation frequency difference, respectively. Based on the analytic hierarchy process, a business priority judgment matrix is ​​constructed, and the eigenvector corresponding to the largest eigenvalue is calculated as the weight value according to the system's tolerance difference for fairness, availability, and bottleneck effect.

[0059] S1.4: Determining the Upgrade Mode of the Hierarchical Model

[0060] Based on longitudinal performance extreme values With lateral performance extremes Two differentiated upgrade modes for the dual-mode communication model were proposed: hard update and soft update.

[0061] Hard update refers to the full joint training of mode selection parameters and channel selection parameters in the model. This mode is suitable for scenarios where the communication environment changes drastically, the system stability collapses, or the mode decision logic fails. It aims to reshape the model structure to quickly restore performance. Soft update refers to locking the mode selection parameters unchanged and only fine-tuning the channel selection parameters. This mode is suitable for scenarios where the mode decision is correct but the local adaptability is insufficient and optimization is required for specific channel details.

[0062] By differentiating between software and hardware updates, a synergistic improvement in computational efficiency and optimization accuracy is achieved. The specific steps are as follows:

[0063] First, regarding the system Calculate the mean value of the longitudinal performance extreme values ​​of each terminal. and variance Construct a comprehensive longitudinal instability index for the system. This index can comprehensively capture the structural instability risk of the system. Vertical weighting. and Based on the system's sensitivity to global structural oscillations and local abnormal fluctuations, normalization is usually used to eliminate the influence of data dimensions and orders of magnitude, thereby achieving a balance between measuring the overall instability trend of the system and the discrete differences of individual terminals.

[0064] Then, statistical analysis of the horizontal extreme value sequence is performed over time to calculate its value within the observation window. mean within and variance Constructing a comprehensive horizontal instability index for the system. , This reflects whether the system is continuously in a state of imbalance or high violation. This reflects the volatility of this imbalance. Horizontal weights. and Depending on the system's sensitivity to the persistence and fluctuations of the imbalance state, normalization is typically used to eliminate the influence of data dimensions and magnitudes in order to achieve a balance between assessing the long-term average violation level and the amplitude of instantaneous performance fluctuations.

[0065] Set the longitudinal stability threshold to The lateral stability threshold is The following judgment rules are set:

[0066] like and At this point, the system is vertically stable and has low horizontal indices. This means there are no drastic oscillations in time, few spatial constraints that are not met, and the differences between iterations are small. This indicates that the current model has good adaptability to the environment and does not require updating.

[0067] like The longitudinal instability index exceeding the limit indicates significant oscillations in the system's temporal and spatial distribution. At this point, the modal decision-making logic is deemed to have failed, necessitating a hard update and full joint training of the mode selection and channel selection parameters.

[0068] like and This indicates that the system is stable longitudinally, but the lateral adaptability index exceeds the limit, indicating that the system violates many constraints or there are large differences between terminals, such as channel competition or mismatch. It is necessary to perform a soft update, count the current optimal mode, lock the mode parameters, and fine-tune the channel parameters only based on the lateral calculation results.

[0069] S2: Upgrade and optimization of the dual-mode communication model based on historical performance gains

[0070] S2.1 Construction of Training Set Based on Historical Sample Gain

[0071] The side server maintains a historical experience pool. This is used to store relevant information generated during past optimization iterations. Each time the system completes a round of performance monitoring and model evaluation and upgrades, it encapsulates the current operating status, decision actions, and evaluation results into an experience sample and stores it in the experience pool. This accumulates environmental evolution and strategy feedback data over long-term operation. (Samples in the experience pool) The empirical tuple is represented as:

[0072] (10)

[0073] The formula is divided into two parts, the first part... The first part contains empirical information used for updating traditional networks. The second part contains additional information. For the sample The modality. For the sample The performance loss vector is represented as ,in, and , and Representing samples respectively Mean and variance of extreme values ​​of longitudinal and lateral performance loss. For the sample The performance gain.

[0074] To make the model focus more on samples that "perform well in similar scenarios," this invention calculates differential sampling weights for each sample. , is represented as:

[0075] (11)

[0076] In the formula, the first term is the performance gain term; a larger value indicates a better learning effect for that sample and a higher sampling weight. The second term is the performance loss matching degree term. Let be the current performance loss vector, denoted as . and , and Let represent the mean and variance of the vertical and horizontal extreme values ​​of the performance loss vector, respectively. Represents the performance loss vector of historical samples With the current performance loss vector The Euclidean distance represents the degree of matching between historical samples and the current terminal's defects and pain points. Samples similar to the current performance loss scenario are selected for learning first. and For the weights of sampling elements, differentiated values ​​are adopted based on different upgrade modes to construct different sample screening logics.

[0077] (1). Soft update mode

[0078] In soft update mode, weight settings limit the selection to samples with the same modality only from the experience pool. At this point, samples with high performance gains and highly similar scenarios are evenly extracted from the experience pool, focusing on local fine-tuning and learning only the fine-tuning under the current modality.

[0079] (2). Hard update mode

[0080] In hard update mode, weights are set and modalities are not limited. High-gain samples are selected from the experience pool, and similarity requirements are relaxed to introduce diverse data to prevent overfitting. Focusing on global reshaping, it is necessary to learn the switching logic between different modalities.

[0081] Based on the modal requirements of the update mode, candidate sub-experience pools that meet the conditions are selected from the experience pool. Based on the differentiated weights, the sub-experience pool is calculated. The sampling weights of the samples are mapped to the probability space using the Softmax function, and then randomly sampled and combined with the current samples to form the training set. .

[0082] S2.2 Model Upgrade and Optimization Based on Masking Mechanism

[0083] A parameter update mask mechanism is introduced to construct a joint edge-side optimization loss function, and the model parameters are updated differentially using a gradient descent method with a mask.

[0084] by The loss function of the two-mode communication model for model parameters Represented as:

[0085] (12)

[0086] In the formula, This indicates the number of samples in the training set. Indicates terminal training set The Middle The performance loss function for each sample.

[0087] Based on the upgrade mode determined in the previous steps, construct the gradient mask vector. ,in and The distribution consists of a mask matrix for mode selection and channel selection parameters, and a matrix where the parameters are associated with the function. The element at the corresponding position is set to 1, otherwise it is set to 0.

[0088] For hard update mode, all parameters need to be updated and settings need to be changed. For soft update mode, when updating channel parameters, the mode parameters are locked and set. .

[0089] The model parameters are iteratively updated using gradient descent on the sides. The update formula explicitly includes a masking operation, expressed as follows:

[0090] (13)

[0091] In the formula, For learning rate, This is the Hadamard product, which is the product of corresponding elements. Loss function for parameters The gradient. These are the parameters of the model updated using gradient descent on the sides.

[0092] After completing the model parameter update for the current round, the system uses real-time data from the original observation window to backtest the performance of the updated model and calculates the performance gain relative to the model before the update. Then, the performance loss vector of the current system is... Modal information With basic status information Forming new experience tuples The data is stored in the historical experience pool of the edge server as prior knowledge for subsequent iterative optimization, thus realizing a closed-loop record of the model evolution process.

[0093] S2.3 Side Model Distribution

[0094] After upgrading and optimizing the dual-mode communication model on the side, the trained model parameters will be... The data is distributed to the corresponding endpoint for synchronous updates of the endpoint model. The update process is represented as follows:

[0095] (14)

[0096] After the model is distributed, the edge device completes local model replacement and activation based on the received parameters, realizing edge-edge collaborative model update.

[0097] This invention achieves joint optimization of a dual-mode communication model by fusing historical performance gains with real-time sample information, significantly improving the convergence efficiency and performance stability of model training. This method can dynamically adjust the optimization range and iteration strategy according to different upgrade modes, ensuring continuous improvement in communication performance while reducing the system's computational burden. This achieves adaptability and efficiency in the model optimization process, enhancing the stability and adaptability of the method in dynamic communication scenarios. Attached Figure Description

[0098] Figure 1 This is a schematic diagram of the dual-mode communication optimization method system of the present invention;

[0099] Figure 2 This is a flowchart of the dual-mode communication optimization method of the present invention. Detailed Implementation

[0100] This invention proposes a dual-mode communication optimization system, such as... Figure 1 As shown, it includes an end-side subsystem and a side-side subsystem;

[0101] The end-side subsystem includes a model synchronization module, a dual-mode communication model decision module, a performance data acquisition module, and a model upload module.

[0102] The side subsystem includes a performance prediction and extreme value analysis module, an upgrade mode judgment module, a model upgrade optimization module based on historical performance gains, and a model distribution module;

[0103] The details are as follows:

[0104] The model synchronization module receives updated dual-mode communication model parameters from the edge and performs parameter replacement and hot loading on the local model. This module ensures that the edge-side decision model is consistent with the edge-side optimization results, and its updated model parameters are input to the dual-mode communication model decision module.

[0105] The dual-mode communication model decision module is used to jointly decide on mode selection and channel selection at the endpoint based on the synchronized dual-mode communication model, given channel state and data priority. This module consists of two parts: a mode selection subnetwork and a channel selection subnetwork. It takes state information as input and outputs the optimal communication mode and channel selection results. Its operational data is input to the performance data acquisition module.

[0106] The performance data acquisition module is used to collect communication performance data from the end-user device in real time during operation and construct a performance time series. This module records raw data such as latency and packet loss rate in each time slot, providing raw data support for subsequent analysis. The acquisition results are input into the model upload module.

[0107] The model upload module is used to establish an uplink data transmission channel from the endpoint to the edge. This module is responsible for packaging and uploading the collected historical performance sequences and the current original dual-mode communication model parameters of the endpoint, and outputting them to the performance prediction and extreme value analysis module and the model upgrade and optimization module based on historical performance gains of the edge subsystem, respectively.

[0108] The performance prediction and extreme value analysis module is used to obtain the performance prediction sequence for future periods by using a rolling prediction mechanism and extracting multi-dimensional features. This module calculates the longitudinal performance extreme value of a single terminal to characterize temporal stability and calculates the lateral performance extreme value of multiple terminals to characterize spatial balance. The output quantitative index is input to the upgrade mode judgment module.

[0109] The upgrade mode determination module is used to determine the system's model upgrade requirements based on statistical methods. This module statistically analyzes the extreme values ​​of each terminal to construct a comprehensive longitudinal instability index and a lateral adaptation index. When the longitudinal instability index exceeds the limit, a structural failure is determined, triggering a hard upgrade mode; when the longitudinal instability index is stable but the lateral adaptation index exceeds the limit, local adaptability is determined, triggering a soft upgrade mode; otherwise, no upgrade is performed. The determination results of this module are input to the model upgrade optimization module based on historical performance gains.

[0110] The model upgrade and optimization module based on historical performance gains receives the original model parameters and performs differentiated model training. This module maintains a historical experience pool and dynamically adjusts sample sampling weights to construct the training set based on the upgrade mode, performance gain, and scene similarity. It also introduces a parameter update mask mechanism: in hard upgrade mode, all parameters are updated; in soft upgrade mode, mode selection parameters are locked through a mask, and only channel selection parameters are optimized. The optimized model parameters are then input to the model distribution module.

[0111] The model distribution module encapsulates the updated model parameters after model optimization and distributes them to the corresponding edge subsystems via a downlink channel. This module transmits parameters through a secure communication channel, ensuring data integrity and model consistency, and enabling data input and closed-loop updates for the edge model synchronization module.

[0112] The dual-mode communication optimization method of this invention comprises two parts: a dual-mode communication model upgrade mode determination method considering long-term performance extrema and a dual-mode communication model upgrade optimization method based on historical performance gains. The flowchart is shown below. Figure 2 As shown.

[0113] S1: Determination of upgrade mode for dual-mode communication model considering long-term performance extremes

[0114] To achieve adaptive upgrades of dual-mode communication systems under dynamic channel environments, this invention proposes a method for determining upgrade modes in dual-mode communication models that considers long-term performance extrema. First, an end-side dual-mode communication model is constructed, and communication links are jointly decided through a mode selection subnetwork and a channel selection subnetwork. Second, a sliding time window mechanism is used to collect performance data and perform rolling predictions to obtain performance sequences for future periods. Then, performance extrema are calculated from both temporal and spatial dimensions. The vertical extrema incorporate a second-order difference term to quantify the oscillation severity of a single terminal, while the horizontal extrema integrate equilibrium and constraint violation frequencies to quantify the system's adaptability defects. Finally, a system-level comprehensive index is constructed based on the mean and variance of each terminal's extrema. An upgrade strategy is adaptively selected through a dual-threshold judgment logic: a hard upgrade is performed to reshape the structure when the system oscillates; a soft upgrade is performed to fine-tune the channel when the system is unbalanced; and the status quo is maintained when the system is stable and balanced. This method achieves accurate hierarchical determination of model upgrade needs, avoiding unnecessary full-scale training overhead while ensuring the stability and robustness of the system in long-term operation.

[0115] S1.1 End-side dual-mode communication model decision

[0116] Considering There are 1 terminal, and the set is denoted as _ . ,terminal Shared with edge servers Channels, represented as a set All channels are divided into two subsets: the set of channels corresponding to the first mode (such as power line carrier). The channel set corresponding to the first and second modes (such as wireless communication) ,and .terminal The decision variables include modal selection variables. and channel selection variables ,when Time indicates terminal Selecting the first communication mode, the terminal from... Select Channel ; Time indicates terminal Selecting the second communication mode, the terminal from Select Channel .

[0117] Each terminal is equipped with a dual-mode communication model. ,in For model parameters, and Representing terminals respectively Mode selection subnetwork parameters and channel selection subnetwork parameters. This is status information, where For channel state, Prioritize the data.

[0118] By input The dual-mode communication model makes a decision and outputs the selected communication mode and the corresponding channel, which is represented as follows;

[0119] (1)

[0120] In the formula, The modality selection result is determined by the parameters of the modality selection subnetwork. The channel selection result is determined by the parameters of the channel selection sub-network.

[0121] S1.2 Performance Data Acquisition and Statistics

[0122] No. Time slot, acquisition terminal latency With packet loss rate Construct a performance loss function , is represented as:

[0123] (2)

[0124] In the formula, and Do not represent the weights of latency and packet loss rate.

[0125] To capture the dynamic evolution of the communication environment, this invention designs a scrolling prediction method based on a sliding time window at the edge. The observation window length is set to... .exist Time slots, for terminals Construct historical performance sequences , is represented as:

[0126] (3)

[0127] Pre-trained prediction model (Models such as LSTM (Long Short-Term Memory) or GRU (Gated Recurrent Unit) are not limited to this invention.) The prediction process employs an iterative prediction mechanism, which involves a "prediction-filling-re-prediction" rolling process. The specific steps are as follows:

[0128] 1. Single-step prediction: using a pre-trained prediction model Predict the output of the next time slot Performance prediction , is represented as:

[0129] (4)

[0130] 2. Rolling Update: The obtained predicted values ​​are updated through a first-in, first-out (FIFO) rolling operation. Replace historical performance sequence The original first data A new historical performance sequence is obtained. And input the prediction model again. Make predictions.

[0131] Repeat the above steps Next, through rolling prediction, a length of [length] is obtained. The prediction performance sequence is represented as:

[0132] (5)

[0133] The sequence It is then output to the performance extreme value comprehensive analysis module for subsequent extreme value feature extraction.

[0134] S1.3 Comprehensive Analysis of Horizontal and Vertical Performance Extremes

[0135] To accurately quantify the system state, this invention defines the longitudinal performance extremum to characterize the temporal stability of a single terminal, and the lateral performance extremum to characterize the spatial balance of multiple terminals.

[0136] Define longitudinal performance extreme values This reflects the degree of performance fluctuation of a single terminal over a future period. Since simple variance cannot reflect the risk of "jumps," this invention introduces a second-order difference term to identify non-stationary states that may lead to frequent mode switching. Specifically, it is expressed as:

[0137] (6)

[0138] In the formula, and The weights are the fluctuation amplitude weight and the oscillation acceleration weight, respectively, which can be determined by calculating the information entropy using the entropy weight method based on historical data. For the terminal The predicted average loss is expressed as The first term is the standard deviation, representing the overall fluctuation of the terminal's performance. The second term is the sum of the absolute values ​​of the second-order differences, representing the acceleration of the change in the terminal's performance. For second-order difference operations, it is represented as:

[0139] (7)

[0140] To address the problem that existing technologies struggle to balance fairness across multiple terminals and performance constraints, this invention constructs a lateral performance extremum from three dimensions: index balance, constraint violation frequency, and constraint violation distribution differences. .

[0141] First, statistical computing terminal In the prediction window Number of times the constraint is not satisfied , is represented as:

[0142] (8)

[0143] In the formula, The performance threshold is determined based on the maximum allowable latency and packet loss rate for the current terminal service type, combined with the aforementioned performance loss function. This is an indicator function; it takes the value 1 if the condition is met, and 0 otherwise.

[0144] Next, calculate Lateral performance extreme values ​​of time slots , is represented as:

[0145] (9)

[0146] In the formula, the first term represents the performance distribution uniformity, reflecting the dispersion of the average performance loss of different terminals during the prediction period. The first term represents the average performance loss across all terminals. The second term represents the system-level constraint violation strength, reflecting the severity of the system's inability to meet performance constraints within a future period. This value normalizes the number of terminals and the time window, directly reflecting the overall unavailability probability of the system. The third term represents the violation frequency variability, reflecting the uneven distribution of violation phenomena across different terminals. The larger this value, the worse the model's adaptability to individual terminals. These are weights for performance distribution balance, system-level constraint violation intensity, and violation frequency difference, respectively. Based on the analytic hierarchy process, a business priority judgment matrix is ​​constructed, and the eigenvector corresponding to the largest eigenvalue is calculated as the weight value according to the system's tolerance difference for fairness, availability, and bottleneck effect.

[0147] S1.4: Determining the Upgrade Mode of the Hierarchical Model

[0148] Based on the above longitudinal performance extreme values With lateral performance extremes This invention proposes two differentiated upgrade modes for a dual-mode communication model: hard update and soft update. Hard update refers to performing full joint training on the mode selection parameters and channel selection parameters in the model. This mode is suitable for scenarios where the communication environment changes drastically, system stability collapses, or mode decision logic fails, aiming to reshape the model structure to quickly restore performance. Soft update refers to locking the mode selection parameters unchanged and only fine-tuning the channel selection parameters. This mode is suitable for scenarios where the mode decision is correct but lacks local adaptability and requires optimization for specific channel details.

[0149] The reason for adopting the above-mentioned hierarchical update mechanism is that traditional model update methods often do not distinguish the severity of environmental changes and uniformly adopt full parameter updates. This "one-size-fits-all" approach not only brings huge computational overhead and resource waste, but also, when only fine-tuning of channel parameters is needed, full updates can easily cause unnecessary oscillations in the modal decision layer, making it difficult for the model to focus and pay attention to the micro-details of channel adjustment. This invention achieves a synergistic improvement in computational efficiency and optimization accuracy by distinguishing between soft and hard updates. The specific steps are as follows:

[0150] First, regarding the system Calculate the mean value of the longitudinal performance extreme values ​​of each terminal. and variance Construct a comprehensive longitudinal instability index for the system. This index can comprehensively capture the structural instability risk of the system. Vertical weighting. and Based on the system's sensitivity to global structural oscillations and local abnormal fluctuations, normalization is usually used to eliminate the influence of data dimensions and orders of magnitude, thereby achieving a balance between measuring the overall instability trend of the system and the discrete differences of individual terminals.

[0151] Then, statistical analysis of the horizontal extreme value sequence is performed over time to calculate its value within the observation window. mean within and variance Constructing a comprehensive horizontal instability index for the system. , This reflects whether the system is continuously in a state of imbalance or high violation. This reflects the volatility of this imbalance. Horizontal weights. and Depending on the system's sensitivity to the persistence and fluctuations of the imbalance state, normalization is typically used to eliminate the influence of data dimensions and magnitudes in order to achieve a balance between assessing the long-term average violation level and the amplitude of instantaneous performance fluctuations.

[0152] Set the longitudinal stability threshold to The lateral stability threshold is The following judgment rules are set:

[0153] like and At this point, the system is vertically stable and has low horizontal indices. This means there are no drastic oscillations in time, few spatial constraints that are not met, and the differences between iterations are small. This indicates that the current model has good adaptability to the environment and does not require updating.

[0154] like The longitudinal instability index exceeding the limit indicates significant oscillations in the system's temporal and spatial distribution. At this point, the modal decision-making logic is deemed to have failed, necessitating a hard update and full joint training of the mode selection and channel selection parameters.

[0155] like and This indicates that the system is stable longitudinally, but the lateral adaptability index exceeds the limit, indicating that the system violates many constraints or there are large differences between terminals, such as channel competition or mismatch. It is necessary to perform a soft update, count the current optimal mode, lock the mode parameters, and fine-tune the channel parameters only based on the lateral calculation results.

[0156] This invention introduces a comprehensive analysis of horizontal and vertical performance extrema to dynamically determine the upgrade requirements of dual-mode communication models, significantly improving the accuracy and foresight of the judgment. This method can adaptively distinguish between hard and soft upgrade modes based on changes in terminal performance, ensuring overall system stability while reducing unnecessary model update frequency. This achieves a synergistic balance between maintaining communication performance and computational overhead, enhancing the method's adaptability and practicality in complex channel environments.

[0157] S2: Upgrade and optimization of the dual-mode communication model based on historical performance gains

[0158] After determining the upgrade mode of the dual-mode communication model, this invention proposes an upgrade optimization method based on historical performance gains to further improve the system's optimization efficiency and relevance in complex channel environments. First, a differentiated sample selection mechanism based on a historical experience pool is constructed. According to the current upgrade mode, sampling weights are dynamically calculated using performance gains and scene similarity. High-value samples are selected from historical records and fused with real-time observation samples to form a training set. Second, a parameter update mask mechanism is introduced to construct a weighted loss function. Differential updates are achieved within a unified optimization framework. In hard upgrade mode, all parameters are jointly trained. In soft upgrade mode, modal parameters are forcibly locked using a mask matrix, and only channel parameters are fine-tuned. Finally, the optimized parameters are sent to the end-user for synchronization. This method can adaptively adjust the optimization strategy according to the severity of environmental changes, converge quickly during performance degradation, and accurately focus during local fine-tuning, thereby achieving efficient evolution of the dual-mode communication model.

[0159] S2.1 Construction of Training Set Based on Historical Sample Gain

[0160] The side server maintains a historical experience pool. This is used to store relevant information generated during past optimization iterations. Each time the system completes a round of performance monitoring and model evaluation and upgrades, it encapsulates the current operating status, decision actions, and evaluation results into an experience sample and stores it in the experience pool. This accumulates environmental evolution and strategy feedback data over long-term operation. (Samples in the experience pool) The empirical tuple is represented as:

[0161] (10)

[0162] The formula is divided into two parts, the first part... The first part contains empirical information used for updating traditional networks. The second part contains additional information. For the sample The modality. For the sample The performance loss vector is represented as ,in, and , and Representing samples respectively Mean and variance of extreme values ​​of longitudinal and lateral performance loss. For the sample The performance gain.

[0163] To make the model focus more on samples that "perform well in similar scenarios," this invention calculates differential sampling weights for each sample. , is represented as:

[0164] (11)

[0165] In the formula, the first term is the performance gain term; a larger value indicates a better learning effect for that sample and a higher sampling weight. The second term is the performance loss matching degree term. Let be the current performance loss vector, denoted as . and , and Let represent the mean and variance of the vertical and horizontal extreme values ​​of the performance loss vector, respectively. Represents the performance loss vector of historical samples With the current performance loss vector The Euclidean distance represents the degree of matching between historical samples and the current terminal's defects and pain points. Samples similar to the current performance loss scenario are selected for learning first. and For the weights of sampling elements, differentiated values ​​are adopted based on different upgrade modes to construct different sample screening logics.

[0166] 1. Soft update mode

[0167] In soft update mode, the weight is set to, for example... Limit the selection to samples with the same modality only from the experience pool. At this point, samples with high performance gains and highly similar scenarios are evenly extracted from the experience pool, focusing on local fine-tuning and learning only the fine-tuning under the current modality.

[0168] 2. Hard update mode

[0169] In hard update mode, the weight is set to, for example... Furthermore, without limiting the modality, high-gain samples are selected from the experience pool, while the similarity requirement is relaxed to introduce diverse data to prevent overfitting and focus on global reshaping. This requires learning the switching logic between different modalities.

[0170] Based on the modal requirements of the update mode, candidate sub-experience pools that meet the conditions are selected from the experience pool. Based on the differentiated weights, the sub-experience pool is calculated. The sampling weights of the samples are mapped to the probability space using the Softmax function, and then randomly sampled and combined with the current samples to form the training set. .

[0171] S2.2 Model Upgrade and Optimization Based on Masking Mechanism

[0172] To achieve differentiated updates of model parameters, this invention introduces a parameter update mask mechanism, constructs a joint side-side optimization loss function, and uses a masked gradient descent method to perform differentiated updates of model parameters.

[0173] by The loss function of the two-mode communication model for model parameters Represented as:

[0174] (12)

[0175] In the formula, This indicates the number of samples in the training set. Indicates terminal training set The Middle The performance loss function for each sample.

[0176] Based on the upgrade mode determined in the previous steps, construct the gradient mask vector. ,in and The distribution consists of a mask matrix for mode selection and channel selection parameters, and a matrix where the parameters are associated with the function. The element at the corresponding position is set to 1, otherwise it is set to 0.

[0177] For hard update mode, all parameters need to be updated and settings need to be changed. For soft update mode, when updating channel parameters, the mode parameters are locked and set. .

[0178] The model parameters are iteratively updated using gradient descent on the sides. The update formula explicitly includes a masking operation, expressed as follows:

[0179] (13)

[0180] In the formula, For learning rate, This is the Hadamard product, which is the product of corresponding elements. Loss function for parameters The gradient. These are the parameters of the model updated using gradient descent on the sides.

[0181] This formula enables the locking of parameters in soft update mode by forcing the gradients corresponding to the modal parameters to zero through a mask; while in hard update mode, the gradients propagate back normally. This mechanism ensures seamless switching between soft and hard updates within the same algorithm framework, reducing the complexity of system implementation while ensuring the strict execution of the update strategy.

[0182] After completing the model parameter update for the current round, the system uses real-time data from the original observation window to backtest the performance of the updated model and calculates the performance gain relative to the model before the update. Then, the performance loss vector of the current system is... Modal information With basic status information Forming new experience tuples The data is stored in the historical experience pool of the edge server as prior knowledge for subsequent iterative optimization, thus realizing a closed-loop record of the model evolution process.

[0183] S2.3 Side Model Distribution

[0184] After upgrading and optimizing the dual-mode communication model on the side, the trained model parameters will be... The data is distributed to the corresponding endpoint for synchronous updates of the endpoint model. The update process is represented as follows:

[0185] (14)

[0186] After the model is distributed, the edge device completes local model replacement and activation based on the received parameters, realizing edge-edge collaborative model update.

[0187] This invention proposes a dual-mode communication model upgrade mode judgment method that takes into account the long-term performance extreme value, realizing accurate hierarchical judgment of model upgrade requirements, which not only avoids unnecessary full training overhead, but also ensures the stability and robustness of the system in long-term operation.

[0188] After determining the upgrade mode of the dual-mode communication model, this invention proposes an upgrade optimization method for the dual-mode communication model based on historical performance gains to further improve the optimization efficiency and targeting of the system in complex channel environments. This method can adaptively adjust the optimization strategy according to the severity of environmental changes, converge quickly when performance degrades, and focus precisely when fine-tuning local areas, thereby achieving efficient evolution of the dual-mode communication model.

Claims

1. A dual-mode communication optimization system, characterized in that, The system includes a model synchronization module, a dual-mode communication model decision module, a performance data acquisition module, and a model upload module deployed on the edge side; and a performance prediction and extreme value analysis module, an upgrade mode judgment module, a model upgrade optimization module based on historical performance gains, and a model distribution module deployed on the edge side. The model synchronization module works in conjunction with the dual-mode communication model decision module to perform hot loading after receiving edge-side parameters and achieve real-time joint decision-making on the edge side through a mode and channel selection sub-network. The performance data acquisition module and the model upload module are responsible for collecting runtime data and establishing uplink channels to upload historical performance sequences and original parameters to the edge side. The performance prediction and extreme value analysis module... The module generates performance sequences for future time periods based on a rolling prediction mechanism and quantifies the temporal longitudinal oscillation stability of a single terminal and the spatial lateral collaborative adaptability of multiple terminals. The upgrade mode judgment module constructs a system-level index based on the statistical characteristics of extreme values ​​to dynamically distinguish between hard update, soft update, and no update modes, thereby accurately determining the model upgrade requirements. The model upgrade optimization module based on historical performance gains utilizes differentiated sample screening and parameter update masking mechanisms to achieve adaptive full reshaping or local fine-tuning of mode and channel parameters. The model distribution module is used to securely distribute the updated model parameters to the end side after optimization, realizing edge-end collaboration and closed-loop adaptive updates.

2. The dual-mode communication optimization system according to claim 1, characterized in that, Specifically, it includes end-side subsystems and edge-side subsystems; The aforementioned end-side subsystem includes a model synchronization module, a dual-mode communication model decision module, a performance data acquisition module, and a model upload module; The model synchronization module is used to receive the updated dual-mode communication model parameters sent from the edge side and perform parameter replacement and hot loading of the local model; this module ensures that the end-side decision model is consistent with the edge-side optimization results, and its output updated model parameters are input to the dual-mode communication model decision module. The dual-mode communication model decision module is used to make joint decisions on mode selection and channel selection at the end side based on the synchronized dual-mode communication model under given channel state and data priority conditions. The module consists of two parts: a mode selection sub-network and a channel selection sub-network. By inputting state information, it outputs the optimal communication mode and channel selection results, and its operation data is input to the performance data acquisition module. The performance data acquisition module is used to collect communication performance data from the end side during operation in real time and construct a performance time series. This module records raw data such as latency and packet loss rate in each time slot, providing raw data support for subsequent analysis. Its collection results are input into the model upload module. The model upload module is used to establish an uplink data transmission channel from the end side to the edge side. This module is responsible for packaging and uploading the collected historical performance sequences and the current original dual-mode communication model parameters of the end side, and outputting them to the performance prediction and extreme value analysis module and the model upgrade and optimization module based on historical performance gains of the edge side subsystem, respectively. The side subsystem includes a performance prediction and extreme value analysis module, an upgrade mode judgment module, a model upgrade optimization module based on historical performance gains, and a model distribution module; The performance prediction and extreme value analysis module is used to obtain the performance prediction sequence for future periods by using the rolling prediction mechanism and extracting multi-dimensional features. This module calculates the longitudinal performance extreme value of a single terminal to characterize the temporal stability and calculates the lateral performance extreme value of multiple terminals to characterize the spatial balance. The output quantitative index is input to the upgrade mode judgment module. The upgrade mode determination module is used to determine the system's model upgrade requirements based on statistical methods. This module statistically analyzes the extreme values ​​of each terminal to construct a comprehensive longitudinal instability index and a lateral adaptability index for the system. When the longitudinal instability index exceeds the limit, the system is deemed to have suffered a structural failure, triggering a hard upgrade mode. When the longitudinal instability index is stable but the lateral adaptability index exceeds the limit, the system is deemed to have insufficient local adaptability, triggering a soft upgrade mode. Otherwise, the upgrade will not be performed; the judgment result of this module is input into the model upgrade and optimization module based on historical performance gains; The model upgrade and optimization module based on historical performance gains is used to receive the original model parameters and perform differentiated model training. This module maintains a historical experience pool and dynamically adjusts the sample sampling weights to construct the training set according to the upgrade mode, performance gain, and scene similarity. At the same time, a parameter update mask mechanism is introduced to update all parameters in hard upgrade mode and lock the mode selection parameters and optimize only the channel selection parameters in soft upgrade mode through a mask. The optimized model parameters are then input into the model distribution module. The model distribution module is used to encapsulate the updated model parameters and distribute them to the corresponding end-side subsystems through the downlink channel after the model optimization is completed. This module transmits parameters through a secure communication channel to ensure data integrity and model consistency, and realizes the data input and closed-loop update of the end-side model synchronization module.

3. A dual-mode communication optimization method, characterized in that, The dual-mode communication optimization system according to claim 1 or 2; the method includes the following steps: S1: Determination of upgrade mode for dual-mode communication model considering long-term performance extremes First, a dual-mode communication model is constructed on the end side, and the communication link is jointly decided by the mode selection subnetwork and the channel selection subnetwork. Second, a sliding time window mechanism is used to collect performance data and perform rolling prediction to obtain the performance sequence for future periods. Then, the performance extreme values ​​are calculated from two dimensions: time domain and spatial domain. The vertical extreme value introduces a second-order difference term to quantify the oscillation severity of a single terminal, and the horizontal extreme value integrates the equilibrium and constraint violation frequency to quantify the system's adaptive defects. Finally, a system-level comprehensive index is constructed based on the mean and variance of the extreme values ​​of each terminal, and an upgrade strategy is adaptively selected through a dual-threshold judgment logic: a hard upgrade is performed to reshape the structure when the system oscillates, a soft upgrade is performed to fine-tune the channel when the system is unbalanced, and the status quo is maintained when the system is stable and balanced. S2: Upgrade and optimization of the dual-mode communication model based on historical performance gains First, a differentiated sample selection mechanism based on a historical experience pool is constructed. According to the current upgrade mode, sampling weights are dynamically calculated using performance gain and scene similarity. High-value samples are selected from historical records and fused with real-time observation samples to form a training set. Second, a parameter update mask mechanism is introduced to construct a weighted loss function. Differentiated updates are achieved under a unified optimization framework. In hard upgrade mode, all parameters are jointly trained. In soft upgrade mode, modal parameters are forcibly locked through a mask matrix, and only channel parameters are finely tuned. Finally, the optimized parameters are sent to the end side to complete synchronization.

4. The dual-mode communication optimization method according to claim 3, characterized in that, S1 specifically includes the following steps: S1.1 End-side dual-mode communication model decision Considering There are 1 terminal, and the set is denoted as _ . ,terminal Shared with edge servers Channels, represented as a set All channels are divided into two subsets: the set of channels corresponding to the first mode. The channel set corresponding to the second mode ,and ;terminal The decision variables include modal selection variables. and channel selection variables ,when Time indicates terminal Selecting the first communication mode, the terminal from... Select Channel ; Time indicates terminal Selecting the second communication mode, the terminal from Select Channel ; Each terminal is equipped with a dual-mode communication model. ,in For model parameters, and Representing terminals respectively Mode selection subnetwork parameters and channel selection subnetwork parameters; This is status information, where For channel state, Prioritize data; By input The dual-mode communication model makes a decision and outputs the selected communication mode and the corresponding channel, which is represented as follows; （1） In the formula, The modality selection result is determined by the parameters of the modality selection subnetwork; The channel selection result is determined by the parameters of the channel selection sub-network. S1.2 Performance Data Acquisition and Statistics No. Time slot, acquisition terminal latency With packet loss rate Construct a performance loss function , is represented as: （2） In the formula, and Do not represent the weights of latency and packet loss rate; Set the observation window length to ;exist Time slots, for terminals Construct historical performance sequences , is represented as: （3） Pre-trained prediction model Or GRU, to perform The prediction process employs a rolling mechanism of "prediction-filling-re-prediction," with the specific steps as follows: (1). Single-step prediction: using a pre-trained prediction model Predict the output of the next time slot Performance prediction , is represented as: （4） (2). Rolling update: The predicted values ​​are updated through a first-in-first-out rolling operation. Replace historical performance sequence The original first data A new historical performance sequence is obtained. And input the prediction model again. Make predictions; Repeat the above steps Next, through rolling prediction, a length of [length] is obtained. The prediction performance sequence is represented as: （5） The sequence It is then output to the performance extreme value comprehensive analysis module for subsequent extreme value feature extraction; S1.3 Comprehensive Analysis of Horizontal and Vertical Performance Extremes Define longitudinal performance extreme values This reflects the degree of performance fluctuation of a single terminal over a future period; a second-order difference term is introduced to identify non-stationary states that may lead to frequent mode switching; specifically expressed as: （6） In the formula, and The weights are fluctuation amplitude weight and oscillation acceleration weight, which can be determined by calculating information entropy using the entropy weight method based on historical data. For the terminal The predicted average loss is expressed as The first term is the standard deviation, representing the overall fluctuation of the terminal's performance; the second term is the sum of the absolute values ​​of the second-order differences, representing the acceleration of the change in the terminal's performance; where... For second-order difference operations, it is represented as: （7） Horizontal performance extreme values ​​are constructed from three dimensions: indicator balance, constraint violation frequency, and constraint violation distribution differences. ; First, statistical computing terminal In the prediction window Number of times the constraint is not satisfied , is represented as: （8） In the formula, The performance threshold is determined based on the maximum allowable latency and packet loss rate for the current terminal service type, combined with the aforementioned performance loss function. This is an indicator function; it takes the value 1 if the condition is met, and 0 otherwise. Next, calculate Lateral performance extreme values ​​of time slots , is represented as: （9） In the formula, the first term represents the performance distribution uniformity, reflecting the dispersion of the average performance loss of different terminals during the prediction period. The first term is the average performance loss across all terminals; the second term is the system-level constraint violation strength, reflecting the severity of the system's inability to meet performance constraints within a certain period. This value normalizes the number of terminals and the time window, directly reflecting the overall unavailability probability of the system; the third term is the violation frequency difference, reflecting the uneven distribution of violation phenomena among different terminals. The larger this value, the worse the model's adaptability to individual terminals. These are weights for performance distribution balance, system-level constraint violation intensity, and violation frequency difference, respectively. Based on the analytic hierarchy process, a business priority judgment matrix is ​​constructed, and the eigenvector corresponding to the largest eigenvalue is calculated as the weight value according to the system's tolerance difference for fairness, availability, and bottleneck effect. S1.4: Determining the Upgrade Mode of the Hierarchical Model Based on longitudinal performance extreme values With lateral performance extremes Two differentiated upgrade modes for the dual-mode communication model were proposed: hard update and soft update. Hard update refers to the full joint training of mode selection parameters and channel selection parameters in the model. This mode is suitable for scenarios where the communication environment changes drastically, the system stability collapses, or the mode decision logic fails. It aims to reshape the model structure to quickly restore performance. Soft update refers to locking the mode selection parameters unchanged and only fine-tuning the channel selection parameters. This mode is suitable for scenarios where the mode decision is correct but the local adaptability is insufficient and optimization is required for specific channel details. By differentiating between software and hardware updates, a synergistic improvement in computational efficiency and optimization accuracy is achieved; the specific steps are as follows: First, regarding the system Calculate the mean value of the longitudinal performance extreme values ​​of each terminal. and variance Construct a comprehensive longitudinal instability index for the system. This index can comprehensively capture the structural instability risk of the system; vertical weighting and Based on the system's sensitivity to global structural oscillations and local abnormal fluctuations, normalization is usually used to eliminate the influence of data dimensions and orders of magnitude, thereby achieving a balance between measuring the overall instability trend of the system and the discrete differences of individual terminals. Then, statistical analysis of the horizontal extreme value sequence is performed over time to calculate its value within the observation window. mean within and variance Construct a comprehensive horizontal instability index for the system. , This reflects whether the system is continuously in a state of imbalance or high violation. This reflects the volatility of this imbalance; horizontal weighting and Based on the system's sensitivity to the persistence and fluctuation of the imbalance state, normalization is usually used to eliminate the influence of the dimensions and magnitude of the data in order to achieve a balance between assessing the long-term average violation level and the instantaneous performance jitter. Set the longitudinal stability threshold to The lateral stability threshold is The following judgment rules are set: like and At this point, the system is vertically stable and has low horizontal indicators; this means that there are no drastic oscillations in time, and there are not many situations that do not meet the constraints in space, and the differences between the orders are not significant; this indicates that the current model has good adaptability to the environment and does not need to be updated. like If the longitudinal instability index exceeds the limit, it indicates that the system has significant oscillations in time and space. At this point, the modal decision logic is deemed to have failed, and a hard update must be performed to conduct full joint training on the modal selection and channel selection parameters. like and This indicates that the system is stable longitudinally, but the lateral adaptability index exceeds the limit, indicating that the system violates many constraints or there are large differences between terminals, such as channel competition or mismatch. It is necessary to perform a soft update, count the current optimal mode, lock the mode parameters, and fine-tune the channel parameters only based on the lateral calculation results.

5. The dual-mode communication optimization method according to claim 4, characterized in that, S2: Upgrade and optimization of the dual-mode communication model based on historical performance gains S2.1 Construction of Training Set Based on Historical Sample Gain The side server maintains a historical experience pool. It is used to store relevant information generated during past optimization iterations; After each round of performance monitoring and model evaluation and upgrade, the system encapsulates the current operating status, decision-making actions, and evaluation results into an experience sample and stores it in the experience pool. This accumulates environmental evolution and strategy feedback data over long-term operation. The empirical tuple is represented as: （10） The formula is divided into two parts, the first part... The first part contains empirical information used for updating traditional networks; the second part contains additional information. For the sample The modality; For the sample The performance loss vector is represented as ,in, and , and Representing samples respectively Mean and variance of extreme values ​​of longitudinal and lateral performance loss; For the sample Performance gains; Calculate differential sampling weights for each sample , is represented as: （11） In the formula, the first term is the performance gain term; a larger value indicates a better learning effect for the sample and a higher sampling weight. The second term is the performance loss matching degree term. Let be the current performance loss vector, denoted as ; and , and Let represent the mean and variance of the vertical and horizontal extreme values ​​of the performance loss vector, respectively; Represents the performance loss vector of historical samples With the current performance loss vector The Euclidean distance represents the matching degree between historical samples and the current terminal's defects and pain points. Samples similar to the current performance loss scenario are selected for learning first. and To determine the weights of the sampling elements, differentiated values ​​are used based on different upgrade modes to construct different sample selection logics; (1). Soft update mode In soft update mode, weight settings limit the selection to samples with the same modality only from the experience pool. At this point, samples with high performance gains and highly similar scenarios are evenly drawn from the experience pool, focusing on local fine-tuning and learning only the fine-tuning under the current modality; (2). Hard update mode In hard update mode, weights are set and modalities are not limited. At this time, high-gain samples are selected from the experience pool, and similarity requirements are relaxed to introduce diverse data to prevent overfitting and focus on global reshaping. It is necessary to learn the switching logic between different modalities. Based on the modal requirements of the update mode, candidate sub-experience pools that meet the conditions are selected from the experience pool. ;Calculate the sub-experience pool based on the differentiated weights. The sampling weights of the samples are mapped to the probability space using the Softmax function, and then randomly sampled and combined with the current samples to form the training set. ; S2.2 Model Upgrade and Optimization Based on Masking Mechanism A parameter update mask mechanism is introduced to construct a joint edge-side optimization loss function, and the model parameters are updated differentially using a gradient descent method with a mask. by The loss function of the two-mode communication model for model parameters Represented as: （12） In the formula, This indicates the number of samples in the training set. Indicates terminal training set The Middle The performance loss function for each sample; Based on the upgrade mode determined in the previous steps, construct the gradient mask vector. ,in and The distribution consists of a mask matrix for mode selection and channel selection parameters, and a matrix where the parameters are associated with the function. The element at the corresponding position is set to 1, otherwise it is set to 0; For hard update mode, all parameters need to be updated and settings need to be changed. For soft update mode, when updating channel parameters, the mode parameters are locked and set. ; The model parameters are iteratively updated using gradient descent on the sides. The update formula explicitly includes a masking operation, expressed as follows: （13） In the formula, For learning rate, This is the Hadamard product, which is the product of corresponding elements. Loss function for parameters The gradient; These are the parameters of the model updated using gradient descent on the edges; After completing the model parameter update for the current round, the system uses real-time data from the original observation window to backtest the performance of the updated model and calculates the performance gain relative to the model before the update. Then, the performance loss vector of the current system is... Modal information With basic status information Forming new experience tuples The data is stored in the historical experience pool of the edge server as prior knowledge for subsequent iterative optimization, thus realizing a closed-loop record of the model evolution process. S2.3 Side Model Distribution After upgrading and optimizing the dual-mode communication model on the side, the trained model parameters will be... The data is distributed to the corresponding endpoint for synchronous updates of the endpoint model. The update process is represented as follows: （14） After the model is distributed, the edge device completes local model replacement and activation based on the received parameters, realizing edge-edge collaborative model update.

Images