Smart converged terminal handover control method and system supporting multiple network access

Abstract

The application discloses a smart fusion terminal switching control method and system supporting multi-network access, and relates to the technical field of communication. The method comprises the following steps: collecting a link state parameter set of a heterogeneous network, a resource state parameter set of a terminal and a service demand characteristic vector of a bearer; inputting the link state parameter into a prediction model to predict a service quality feasible region of a candidate network in a future time window; generating an active switching trigger signal if a service quality feasible region margin of a current network is lower than a threshold value; deciding an optimal target network identifier and an optimal switching execution time; issuing a hardware preparation instruction set to the terminal according to the target network identifier before the execution time arrives; and performing an atomic data flow table migration operation to switch a data flow outlet atom of the current network to the target network at the execution time. The application solves the problem of insufficient switching decision timeliness and reliability of existing terminal network switching control, and improves the switching decision timeliness and reliability.

Description

Technical Field

[0001] This application relates to the field of communication technology, and in particular to a method and system for switching control of intelligent converged terminals that supports multiple network access. Background Technology

[0002] When intelligent converged terminals access multiple heterogeneous networks, ensuring the continuity of service quality and transmission stability of the currently carried services is one of the core challenges. Current common technologies for addressing this issue primarily employ handover mechanisms triggered by a single signal strength threshold or fixed rules. Specifically, the terminal periodically measures basic link parameters such as the received signal strength and signal-to-noise ratio of each network. When the signal strength of the currently serving network falls below a preset threshold and the signal strength of a candidate network rises above another threshold, network handover is triggered. However, these existing methods rely solely on instantaneous or very short-term physical layer signal indicators, failing to perceive the dynamic changes in network status over future time windows. This leads to frequent false handovers or delayed handovers in scenarios with frequent signal fluctuations, resulting in service data interruptions or packet loss.

[0003] At present, the network switching control of intelligent converged terminals suffers from technical problems such as insufficient timeliness and reliability of switching decisions. Summary of the Invention

[0004] This application provides a handover control method and system for intelligent converged terminals that supports multi-network access. By real-time collection of link status, terminal resource status, and current service quality requirements of multiple heterogeneous networks, and utilizing an online incremental learning prediction model based on time series, it predicts the feasible service quality domain of each candidate network within a future time window and calculates the feasible domain margin of the current service network. When the margin falls below a dynamically set safety threshold, it triggers an active handover. Upon responding to the trigger signal, it comprehensively considers link status, resource status, service requirements, and prediction results to determine the optimal target network and handover execution time. Before the handover time arrives, it issues a pre-preparation instruction to the hardware abstract scheduling layer. At the handover time, it atomically migrates the data flow table, seamlessly switching the data flow exit to the target network and maintaining the continuity of upper-layer service connections. These technical means solve the technical problems of insufficient timeliness and reliability in the network handover control of existing intelligent converged terminals, achieving the technical effect of improving the timeliness and reliability of handover decisions.

[0005] This application provides a handover control method for intelligent converged terminals supporting multi-network access, comprising: real-time acquisition of link state parameter sets of at least two heterogeneous networks accessed by the intelligent converged terminal, resource state parameter sets of the terminal itself, and service demand feature vectors of the currently carried services; inputting the link state parameter sets into an online incremental learning prediction model based on time series to predict the end-to-end service quality feasible domain of each candidate network within a future time window, and determining whether the service quality feasible domain margin of the current service network is lower than a preset safety threshold based on the service demand feature vectors; if lower, generating an active handover trigger signal; responding to the active handover trigger signal, determining the optimal target network identifier and optimal handover execution time based on the link state parameters, resource state parameters, service demand feature vectors, and service quality feasible domain prediction results; before the optimal handover execution time arrives, issuing a hardware pre-preparation instruction set to the hardware abstraction scheduling layer of the intelligent converged terminal based on the optimal target network identifier; and at the optimal handover execution time, performing an atomic-level data flow table migration operation to atomically switch the data flow exit of the current service network to the optimal target network while maintaining the continuity of upper-layer service connections.

[0006] In a possible implementation, the following processing is performed: the hardware pre-preparation instruction set includes: wake-up and parameter pre-configuration of the target communication module, pre-establishment of the wireless bearer, and pre-allocation of the data buffer; the hardware pre-preparation instruction set also includes an instruction to start a double-buffered rolling storage mechanism, wherein the data to be sent from the current service network is simultaneously written into the buffer pre-allocated to the target network, and the data consistency verification between the two buffers is maintained. After the atomic handover is completed, the data is immediately resumed from the buffer of the target network, achieving zero loss of service data during the handover process.

[0007] In a possible implementation, the following processing is performed: The service requirement feature vector of the current service is obtained by capturing the transport layer protocol type, port number, packet arrival interval and packet length distribution of the current data stream as raw feature parameters; inputting the raw feature parameters into a preset service classifier and outputting a service type label; querying a preset service quality constraint mapping table based on the service type label to obtain the maximum latency tolerance, minimum bandwidth requirement, maximum packet loss tolerance and service priority, which together constitute the service requirement feature vector.

[0008] In a possible implementation, the link state parameter set is input into an online incremental learning prediction model based on time series to predict the end-to-end service quality feasible region for each candidate network within a future time window. The following processing is performed: the link state parameter set is input into the online incremental learning prediction model based on time series, and a confidence interval is added to each predicted value based on the statistical distribution of historical prediction errors. The upper and lower bounds of the confidence interval constitute the feasible region boundary of the corresponding network in the service quality indicator dimension. When the service demand constraints within the future time window completely fall within the feasible region boundary, it is determined that the corresponding network has service quality feasibility within this time window.

[0009] In a possible implementation, the following processing is performed: the parameters of the online incremental learning prediction model are updated online based on historical prediction errors, including: obtaining actual measured values ​​of link states, comparing the actual measured values ​​with the model prediction values, and calculating the prediction error; using the prediction error to correct the model parameters through a recursive optimization algorithm, so that the model's response to changes in network state lags by no more than one prediction period.

[0010] In a possible implementation, in response to the active handover trigger signal, based on the link state parameters, resource state parameters, service demand feature vector, and service quality feasible region prediction results, the optimal target network identifier and optimal handover execution time are determined, and the following processing is performed: the link state parameters, resource state parameters, service demand feature vector, and service quality feasible region prediction results are used to construct a state space; the handover comprehensive benefit function is used as the reward signal; a decision engine based on deep reinforcement learning is invoked to output the optimal target network identifier and optimal handover execution time; wherein, the handover comprehensive benefit function integrates bandwidth gain, latency improvement, power consumption reduction, handover overhead, and ping-pong handover suppression factor, and the weight of each factor is dynamically adjusted according to the service demand feature vector.

[0011] In possible implementations, the optimal target network identifier and optimal switching execution time are determined by performing the following processes: A search time axis is constructed using the current time as the starting point and a future preset time window as the ending point, with a discretized time step; for each candidate network, the service quality feasible domain prediction results are compared with the business demand feature vector, and network-time pairs that satisfy business constraints at each time point on the time axis are selected to form a feasible boundary space; within the feasible boundary space, policy network inference is performed with the switching comprehensive benefit function as the objective, to solve for the network identifier that maximizes the benefit and the corresponding switching execution time; when the feasible boundary space is empty, the decision engine outputs the current service network as the target network and outputs a warning flag, while simultaneously triggering an emergency handling process.

[0012] In a possible implementation, the following processing is performed: the switching comprehensive benefit function considers the expected dwell time after switching during the search process. If a candidate network is expected to maintain a service quality feasible domain exceeding a preset long-term threshold after switching in, an additional long-term stability reward is added to the benefit function; if it is predicted that another switch will be needed shortly after switching in, a discount penalty for future switching costs is added.

[0013] In a possible implementation, the data flow exit atomic of the current service network is switched to the optimal target network while maintaining the continuity of the upper-layer service connection. The following processing is performed: In the data flow forwarding table, the output interface identifier atomic associated with the current data flow is replaced with the output interface identifier of the optimal target network, while keeping the network layer address, transport layer port and encryption context of the data flow unchanged. The replacement operation is completed within an uninterrupted instruction cycle.

[0014] This application also provides a smart converged terminal handover control system supporting multi-network access, including: a status parameter acquisition module, used to collect in real time the link status parameter sets of at least two heterogeneous networks accessed by the smart converged terminal, the resource status parameter set of the terminal itself, and the service demand feature vector of the currently carried service; and an active handover trigger signal generation module, used to input the link status parameter sets into an online incremental learning prediction model based on time series, predict the end-to-end service quality feasible region of each candidate network within a future time window, and determine whether the service quality feasible region margin of the current service network is lower than a preset safety threshold based on the service demand feature vector. If it is lower, an active handover trigger signal is generated. The system includes: a decision module, which responds to the active handover trigger signal and, based on the link status parameters, resource status parameters, service demand feature vectors, and service quality feasible domain prediction results, determines the optimal target network identifier and the optimal handover execution time; a hardware pre-preparation instruction set issuance module, which, before the optimal handover execution time arrives, issues a hardware pre-preparation instruction set to the hardware abstraction scheduling layer of the intelligent converged terminal according to the optimal target network identifier; and a network handover module, which, at the optimal handover execution time, performs an atomic-level data flow table migration operation to atomically switch the data flow exit of the current service network to the optimal target network while maintaining the continuity of upper-layer service connections.

[0015] The proposed intelligent converged terminal handover control method and system supporting multi-network access, as described in this application, firstly collects in real-time the link state parameter sets of at least two heterogeneous networks accessed by the intelligent converged terminal, the terminal's own resource state parameter set, and the service demand feature vector of the currently carried services. Next, the link state parameter sets are input into an online incremental learning prediction model based on time series to predict the end-to-end service quality feasible domain for each candidate network within a future time window. Based on the service demand feature vector, it is determined whether the service quality feasible domain margin of the current serving network is lower than a preset safety threshold. If it is, an active handover trigger signal is generated. Then, in response to the active handover trigger signal, based on the link state parameters, resource state parameters, service demand feature vector, and service quality feasible domain prediction results, the optimal target network identifier and optimal handover execution time are determined. Before the optimal handover execution time arrives, a hardware pre-preparation instruction set is issued to the hardware abstraction scheduling layer of the intelligent converged terminal according to the optimal target network identifier. Finally, at the optimal handover execution time, an atomic-level data flow table migration operation is performed to atomically switch the data flow exit of the current serving network to the optimal target network, while maintaining the continuity of upper-layer service connections. The method and system proposed in this application achieve the technical effect of improving the timeliness and reliability of switching decisions. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments of the present invention will be briefly described below. Flowcharts are used in this application to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed precisely in sequence. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from these processes.

[0017] Figure 1 This is a flowchart illustrating the intelligent converged terminal switching control method supporting multiple network access provided in an embodiment of this application.

[0018] Figure 2 This is a schematic diagram of the structure of an intelligent converged terminal switching control system supporting multiple network access, provided in an embodiment of this application.

[0019] Figure labeling: Status parameter acquisition module 10, active switching trigger signal generation module 20, decision module 30, hardware pre-preparation instruction set issuance module 40, network switching module 50. Detailed Implementation

[0020] The above description is merely an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, specific embodiments of this application are given below.

[0021] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description of this application will be provided in conjunction with the accompanying drawings. The described embodiments should not be considered as limitations on this application. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0022] In the following description, references to "some embodiments" describe a subset of all possible embodiments; however, it is understood that "some embodiments" may be the same or different subsets of all possible embodiments and may be combined with each other without conflict. The terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or modules not explicitly listed or inherent to such processes, methods, products, or devices. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only.

[0023] This application provides a method for switching control of intelligent converged terminals that supports multiple network accesses, such as... Figure 1 As shown, the method includes:

[0024] Step S100: Real-time collection of link status parameter sets of at least two heterogeneous networks accessed by the intelligent converged terminal, resource status parameter sets of the terminal itself, and service requirement feature vectors of the services currently carried.

[0025] Specifically, the heterogeneous network management module built into the intelligent converged terminal actively reads the link status parameters of each connected network according to a preset sampling period by calling the driver interfaces of various communication modules, including but not limited to 5G modules, Wi-Fi modules, and Ethernet modules. This link status parameter set includes signal received strength, signal-to-noise ratio, available bandwidth, round-trip time, packet loss rate, and network load rate. Simultaneously, the terminal's own resource status parameter set is provided by the terminal's operating system resource monitoring service, including current CPU utilization, remaining memory capacity, remaining battery power, and temperature. The service requirement feature vector of the currently carried service is generated by the service awareness module.

[0026] In one possible implementation, obtaining the service requirement feature vector of the currently carried service, step S100 further includes step S110, capturing the transport layer protocol type, port number, packet arrival interval, and packet length distribution of the current data stream as raw feature parameters. Specifically, the service awareness module intercepts the packet header of the currently transmitted data stream through hook functions in the kernel network protocol stack, such as the NF_IP_PRE_ROUTING hook point in the Linux Netfilter framework, and extracts the transport layer protocol field, source port number, and destination port number from the header. Simultaneously, the service awareness module calls a high-precision timestamp counter to record the arrival times of two consecutive data packets, calculates the arrival interval of adjacent data packets, and stores 100 consecutive data packet arrival intervals as an array; for each data packet, it obtains the data packet length field from the network card driver layer, records the packet length distribution, and generates a packet length histogram.

[0027] Step S120: Input the original feature parameters into a preset service classifier and output service type labels. Specifically, the service classifier is an offline pre-trained model based on a decision tree, trained using the C4.5 algorithm. Its input feature vector dimensions are [transport layer protocol type, destination port number, mean packet arrival interval, variance of packet arrival interval, mean packet length, variance of packet length]. During deployment, the decision tree model is embedded in the terminal's read-only memory as executable code. During runtime, the service awareness module concatenates the original feature parameters collected in step S110 into a feature vector and inputs it into the decision tree model. Starting from the root node, the model compares the feature values ​​with the node thresholds layer by layer, and finally outputs the service type labels corresponding to the leaf nodes, such as voice calls, video streams, file downloads, instant messaging, or web browsing.

[0028] Step S130: A preset service quality constraint mapping table is queried based on the service type label to obtain the maximum latency tolerance, minimum bandwidth requirement, maximum packet loss tolerance, and service priority, which together constitute the service requirement feature vector. Specifically, the service quality constraint mapping table is stored in the terminal's non-volatile memory in key-value pair format, where the key is the service type label and the value is a quadruple [maximum latency tolerance, minimum bandwidth requirement, maximum packet loss tolerance, service priority]. For example, the mapping record corresponding to the service type label "voice call" is [maximum latency tolerance = 100ms, minimum bandwidth requirement = 64kbit / s, maximum packet loss tolerance = 1%, service priority = 6], where the service priority ranges from 0 to 7, with higher values ​​indicating higher priority. After reading this quadruple, the service awareness module outputs it as the service requirement feature vector.

[0029] Step S200: Input the link state parameter set into the online incremental learning prediction model based on time series to predict the end-to-end service quality feasible domain of each candidate network within the future time window, and determine whether the service quality feasible domain margin of the current service network is lower than the preset safety threshold based on the service demand feature vector. If it is lower, generate an active handover trigger signal.

[0030] Specifically, the handover control module of the intelligent converged terminal reads the set of link state parameters collected in step S100, arranged in chronological order, from the data buffer at fixed time intervals. This handover control module integrates an online incremental learning prediction model based on time series. This model can use historical link state data from the current and past times to predict the possible range of multiple service quality indicators for each candidate network within a future period. Due to the random fluctuations in the wireless network environment, the prediction result is not a fixed value, but a feasible range defined by upper and lower bounds of a confidence interval. This range is defined as the end-to-end service quality feasible domain of the candidate network within the future time window. To ensure prediction accuracy, the model operates using an online incremental learning method. That is, whenever a new actual measurement value of the link state is obtained, the model will adjust its parameters in real time based on the prediction error, thereby quickly tracking changes in network state. After obtaining the feasible service quality domain (SQV) for each candidate network within a future time window, the handover control module compares the boundary of the current service network's feasible domain with the service demand feature vector obtained in step S100, and calculates the percentage of the SQV margin that the current service network can still provide relative to the service demand. This percentage is the SQV feasible domain margin. Simultaneously, the handover control module dynamically calculates a safety threshold for triggering a handover based on the priority of the current service and the strictness of the SQV constraints, combined with the historical handover success rate and the stability of the current network. If the calculated SQV feasible domain margin is lower than this dynamic safety threshold, it is determined that the current service network will soon be unable to meet the service quality requirements of the service. The handover control module immediately generates an active handover trigger signal and sends it to the decision engine in step S300; otherwise, no trigger signal is generated, and monitoring continues into the next cycle. The dynamic setting rules for the preset security threshold are as follows: for services with high priority and extremely low latency tolerance, the security threshold is set higher to trigger the handover in advance; for non-real-time or high-tolerance services, the security threshold is set lower to avoid frequent handovers. At the same time, the higher the historical handover success rate or the more stable the current network, the higher the security threshold, and vice versa.

[0031] In one possible implementation, the link state parameter set is input into an online incremental learning prediction model based on time series to predict the end-to-end service quality feasible region for each candidate network within a future time window. Step S200 further includes step S210, where the link state parameter set is input into the online incremental learning prediction model based on time series, and a confidence interval is added to each predicted value based on the statistical distribution of historical prediction errors. The upper and lower bounds of the confidence interval constitute the feasible region boundary of the corresponding network in the service quality index dimension. Specifically, the online incremental learning prediction model adopts an online autoregressive integral moving average model, with a fixed model order of 5 for autoregression, 1 for differencing, and 2 for moving average. The link state parameters collected at each time t are used to form a time series numerical sequence, and the model outputs the predicted value for each time within a future time window in each prediction period. Simultaneously, a sliding window of length 20 is maintained to store the prediction error between the most recent 20 predicted values ​​and the actual measured values. Calculate the mean and standard deviation of the prediction error within the sliding window. The upper limit of the confidence interval is equal to the predicted value plus one standard deviation, and the lower limit of the confidence interval is equal to the predicted value minus one standard deviation. For each candidate network, extract the upper and lower bounds of the confidence intervals for its three quality of service indicators: available bandwidth, round-trip time, and packet loss rate. These together constitute the three-dimensional feasible domain boundary of the network.

[0032] Step S220: When the service demand constraints within the future time window completely fall within the feasible domain boundary, the corresponding network is determined to have service quality feasibility within the current time window. Specifically, the switching control module extracts the maximum latency tolerance, minimum bandwidth requirement, and maximum packet loss tolerance from the service demand feature vector output in step S130, and compares them with the feasible domain boundary of each candidate network output in step S210. The determination condition is that all three of the following conditions are met simultaneously: the minimum bandwidth requirement in the service demand feature vector is greater than or equal to the lower limit of the network bandwidth confidence interval; the maximum latency tolerance in the service demand feature vector is less than or equal to the upper limit of the network round-trip delay confidence interval; and the maximum packet loss tolerance in the service demand feature vector is greater than or equal to the upper limit of the network packet loss rate confidence interval. If all three conditions are met simultaneously, the network is determined to have service quality feasibility within the corresponding future time window; otherwise, it is determined not to have it.

[0033] In one possible implementation, the parameters of the online incremental learning prediction model are updated online based on historical prediction errors. Step S210 further includes step S211, which involves obtaining the actual measured value of the link state, comparing the actual measured value with the model prediction value, and calculating the prediction error. Specifically, after each sampling cycle, the hardware abstraction layer obtains the actual measured value of the link state from the communication module driver, reads the prediction value corresponding to the previous prediction cycle from the model cache, subtracts the prediction value from the actual measured value to obtain the prediction error at that moment, and pushes the prediction error into the sliding window described in step S210.

[0034] Step S212: The model parameters are corrected using a recursive optimization algorithm based on the prediction error, ensuring that the model's response lag to network state changes does not exceed one prediction cycle. Specifically, the recursive optimization algorithm is a recursive least squares algorithm, with a forgetting factor set to 0.98. The algorithm execution flow is as follows: First, the online autoregressive integral moving average model is converted into an equivalent autoregressive moving average model form, obtaining a parameter vector. The recursive least squares algorithm maintains a gain vector and a covariance matrix. For each prediction error, the parameter correction amount is calculated according to the standard recursive formula of the recursive least squares algorithm, and the original parameter vector is added to the correction amount to obtain the updated parameter vector. The updated parameters are written back to the model for prediction at the next time step. Multiple parameter updates are performed within each prediction cycle, and the parameter update frequency is not lower than the sampling frequency of the prediction cycle, enabling the model to complete parameter correction based on the latest network state before the end of the current prediction cycle, thereby ensuring that the model's response lag to network mutations does not exceed one prediction cycle.

[0035] Step S300: In response to the active handover trigger signal, based on the link status parameters, resource status parameters, service demand feature vector, and service quality feasible domain prediction results, determine the optimal target network identifier and the optimal handover execution time.

[0036] Specifically, upon receiving the active handover trigger signal generated in step S200, the handover control module uses this signal as the interrupt source for decision initiation and immediately invokes its internal decision engine. The decision engine first reads the current link state parameter set, the terminal's own resource state parameter set, the service demand feature vector generated in step S130, and the service quality feasible region prediction results for each candidate network at each time point within the future time window, all from the data buffer. These data together constitute the state space required for decision-making. The decision engine internally contains a solver with a handover comprehensive benefit function as the optimization objective. This solver is responsible for selecting the network identifier and corresponding execution time that maximizes the comprehensive benefit from all combinations of candidate networks and future selectable handover times, while satisfying service quality constraints. The solution process includes: first, eliminating network-time combinations that cannot meet business requirements based on the service quality feasible domain prediction results; then, calculating the comprehensive switching benefits for each remaining feasible combination. This benefit comprehensively considers the positive gains such as bandwidth increase, latency improvement, and power consumption reduction after switching, while deducting the overhead of the switching operation itself and the risk cost of potential ping-pong switching. The weights of each benefit and cost are dynamically adjusted according to the characteristics of the current business. The decision engine outputs two results: one is the optimal target network identifier, indicating the heterogeneous network to be accessed after switching; the other is the optimal switching execution time, indicating at which specific time in the future the atomic-level data flow table migration operation will be performed. If no network-time pair satisfying the business constraints is found among the feasible combinations, the decision engine outputs the current service network as the target network, outputs a warning flag, and triggers an emergency handling process, sending a service quality degradation notification to the application layer.

[0037] In one possible implementation, in response to the active handover trigger signal, based on the link state parameters, resource state parameters, service demand feature vector, and service quality feasible region prediction results, the optimal target network identifier and optimal handover execution time are determined. Step S300 further includes step S310, where the link state parameters, resource state parameters, service demand feature vector, and service quality feasible region prediction results together constitute a state space, and a deep reinforcement learning-based decision engine is invoked using a handover comprehensive benefit function as a reward signal to output the optimal target network identifier and optimal handover execution time; wherein, the handover comprehensive benefit function integrates bandwidth gain, latency improvement, power consumption reduction, handover overhead, and ping-pong handover suppression factor, and the weights of each factor are dynamically adjusted according to the service demand feature vector.

[0038] Specifically, the decision engine internally deploys a deep reinforcement learning-based decision model, which is pre-trained using a dual-delay deep deterministic policy gradient algorithm. The decision engine reads the link state parameters of all candidate networks at the current time, the terminal's own resource state parameters, the service demand feature vector output in step S130, and the service quality feasible region prediction results for each candidate network at each time step within a future time window, all from the data buffer. This data is then concatenated into a multi-dimensional vector according to a preset dimensional order; this multi-dimensional vector constitutes the state space. The decision engine inputs this state space vector into the policy network of the deep reinforcement learning model. The policy network outputs an action vector, which, after decoding, yields the optimal target network identifier and the optimal handover execution time. The decision-making process uses the switching comprehensive benefit function as a reward signal to guide the action selection of the policy network. The switching comprehensive benefit function is calculated as follows: First, the bandwidth gain is calculated, which is equal to the predicted available bandwidth of the target network minus the available bandwidth of the current serving network; second, the latency improvement is calculated, which is equal to the round-trip latency of the current serving network minus the round-trip latency of the target network; third, the power consumption reduction is calculated, which is equal to the power consumption of the corresponding communication module of the current serving network minus the power consumption of the corresponding communication module of the target network. The power consumption values ​​of each module are provided by the hardware abstraction layer; the switching overhead is taken as a fixed constant value; the ping-pong switching suppression factor is used to reflect the frequency of recent switching. Its value is negatively correlated with the number of switching that occurred within the preset time window in the past. That is, the more switching there are, the smaller the value of this factor is, thereby reducing the comprehensive benefit in the frequent switching scenario and suppressing the ping-pong effect. For example, the ping-pong switching suppression factor can be taken as a power of n of a constant less than 1, where n is the number of switching that occurred in the past period. The weights of each factor are dynamically adjusted based on the service priority in the service demand feature vector. For example, when the service priority is greater than or equal to 5, the weight of latency improvement is set to 0.4, bandwidth gain to 0.3, power consumption reduction to 0.1, handover overhead to 0.1, and ping-pong handover suppression factor to 0.1. When the service priority is less than 5, the weights are: bandwidth gain to 0.4, latency improvement to 0.2, power consumption reduction to 0.2, handover overhead to 0.1, and ping-pong handover suppression factor to 0.1. The overall handover benefit equals bandwidth gain multiplied by its weight, plus latency improvement multiplied by its weight, plus power consumption reduction multiplied by its weight, minus handover overhead multiplied by its weight, and finally minus ping-pong handover suppression factor multiplied by its weight.

[0039] In one possible implementation, the optimal target network identifier and optimal switching execution time are determined. Step S310 further includes step S311, which constructs a search time axis with the current time as the starting point and a future preset time window as the ending point, using a discretized time step. Specifically, the decision engine obtains the current time from the system clock as the starting point of the time axis and uses the future time window length output by the prediction model in step S210 as the ending point of the time axis. The decision engine divides the time period between the starting point and the ending point according to a fixed discretized time step, thereby constructing a search time axis containing multiple discrete time points.

[0040] Step S312: For each candidate network, the predicted service quality feasible region is compared with the service demand feature vector to select network-time pairs that satisfy the service constraints at all times on the time axis, forming a feasible boundary space. Specifically, the decision engine traverses each candidate network. For each discrete time point on the search time axis constructed in step S311, the service quality feasible region boundary predicted in step S210 for that network at that time is extracted, including the lower limit of the available bandwidth confidence interval, the upper limit of the round-trip delay confidence interval, and the upper limit of the packet loss rate confidence interval. Simultaneously, the minimum bandwidth requirement, maximum latency tolerance, and maximum packet loss tolerance are extracted from the service demand feature vector output in step S130. The decision engine performs three comparisons: determining whether the minimum bandwidth requirement is less than or equal to the lower limit of the available bandwidth confidence interval, determining whether the maximum latency tolerance is greater than or equal to the upper limit of the round-trip delay confidence interval, and determining whether the maximum packet loss tolerance is greater than or equal to the upper limit of the packet loss rate confidence interval. Only when all three comparison results are true is the network determined to satisfy the service constraints at that time. For each network-time pair that satisfies the business constraints, i.e., a combination of a specific candidate network identifier and a specific discrete time point, the decision engine records it as a feasible entry, and the set of all feasible entries constitutes the feasible boundary space.

[0041] Step S313: Within the feasible boundary space, with the switching comprehensive benefit function as the objective, policy network inference is performed to solve for the network identifier that maximizes the benefit and the corresponding switching execution time. When the feasible boundary space is empty, the decision engine outputs the current service network as the target network and outputs a warning flag, simultaneously triggering an emergency handling process. Specifically, the decision engine first checks whether the feasible boundary space generated in step S312 is empty. If the feasible boundary space is not empty, the decision engine inputs the state space vector constructed in step S310 into the policy network of the deep reinforcement learning model for forward inference. The policy network outputs an action vector, the dimension of which is equal to the number of candidate networks multiplied by the number of discrete time points on the search time axis. The decision engine performs softmax normalization on the action vector, converting it into a probability distribution, and then selects the network-time pair corresponding to the action dimension with the highest probability value as the optimal decision result. The network identifier part is output as the optimal target network identifier, and the time part is output as the optimal switching execution time. If the feasible boundary space is empty, then no network can satisfy the business constraints at any time within the future time window. The decision engine outputs the identifier of the current service network as the target network, generates a warning flag and sets it to 1, and triggers an emergency handling process. This emergency handling process includes: calling the operating system's notification interface to send a service quality degradation message to the application layer, notifying the upper-layer application that there is currently no available network that meets the business requirements; raising the priority of dropping the current business data packets in the network protocol stack to the highest level; and sending a bitrate reduction request to the business awareness module, suggesting that the upper-layer application reduce the current business's sending bitrate to adapt to the current network capabilities.

[0042] In one possible implementation, step S313 further includes step S3131, wherein the switching comprehensive benefit function considers the expected dwell time after switching during the search process. If a candidate network is expected to maintain a service quality feasible region for more than a preset long time threshold after switching in, an additional long-term stability reward is added to the benefit function; if it is predicted that another switch will be needed in a short period of time after switching in, a discount penalty for future switching costs is added.

[0043] Specifically, when calculating the overall handover benefit for each candidate network in step S313, the decision engine introduces an additional correction term related to the expected dwell time after handover. First, based on the service quality feasible region prediction results output in step S210, the decision engine counts the number of discrete time points at which each candidate network can continuously satisfy business constraints after the handover. This number is multiplied by the discretized time step to obtain the expected dwell time. If the expected dwell time of a candidate network exceeds a preset long-term threshold, an additional long-term stability reward value is added to the overall handover benefit of that network. This reward value can be a fixed constant. Simultaneously, the decision engine looks backward on the search timeline, checking whether there are multiple consecutive discrete time points at which the candidate network fails to meet business constraints within the time period after the expected dwell time after handover. If so, it is determined that the network needs to be handovered again shortly after handover. In this case, a discount penalty for future handover costs is added to the overall handover benefit of the network. This penalty is calculated by first calculating the standard handover cost, then multiplying the standard handover cost by a discount factor to obtain the discount penalty for future handover costs. The final overall benefit of switching used for decision-making is equal to the basic overall benefit calculated in step S310 plus the long-term stability reward, or minus the discount penalty for future switching costs.

[0044] Step S400: Before the optimal handover execution time arrives, a hardware pre-preparation instruction set is issued to the hardware abstraction scheduling layer of the intelligent converged terminal according to the optimal target network identifier. The hardware pre-preparation instruction set includes: wake-up and parameter pre-configuration of the target communication module, pre-establishment of the wireless bearer, and pre-allocation of the data buffer. The hardware pre-preparation instruction set also includes an instruction to start a dual-buffer rolling storage mechanism, wherein the data to be sent from the current service network is simultaneously written into the buffer pre-allocated to the target network, and the data consistency verification of the two buffers is maintained. After the atomic handover is completed, the data is immediately resumed from the buffer of the target network to achieve zero loss of service data during the handover process.

[0045] Specifically, after the decision engine outputs the optimal target network identifier and optimal handover execution time in step S313, it writes the optimal handover execution time into a hardware timer register. This hardware timer is configured to generate an interrupt signal when the optimal handover execution time is reached. Simultaneously, the decision engine generates a hardware pre-preparation instruction set based on the optimal target network identifier and sends this instruction set to the hardware abstraction scheduling layer via the kernel-mode system call interface. Upon receiving the instruction set, the hardware abstraction scheduling layer first parses the optimal target network identifier and performs the corresponding pre-preparation operation based on the network type corresponding to that identifier. For example, if the optimal target network identifier corresponds to a cellular network such as a 5G network, the hardware abstraction scheduling layer calls the power management interface of the corresponding cellular module, sends a wake-up signal to the module's wake-up pin, switches the module from deep sleep to wake-up, and then sends parameters such as the access point name, bearer identifier, and quality of service flow identifier to the module for pre-configuration using the instruction set supported by the module, such as serial AT commands. Next, it initiates a pre-establishment process for the radio bearer with the base station via control plane signaling, including sending a protocol data unit session establishment request and waiting for a response confirmation. If the optimal target network identifier corresponds to a wireless LAN such as Wi-Fi, the hardware abstraction scheduling layer calls the driver interface of the corresponding wireless LAN module to scan the specified service set identifier and trigger the authentication and association process. For other types of networks such as Ethernet, the corresponding link layer pre-configuration operation is performed. The hardware pre-preparation instruction set also includes an instruction to start a double-buffered rolling storage mechanism. The hardware abstraction scheduling layer executes this instruction as follows: it allocates two circular buffers of equal size in the terminal system memory. The first buffer is bound to the sending queue of the current service network, and the second buffer is bound to the pre-allocated buffer of the optimal target network. When each data packet is sent, the hardware abstraction scheduling layer copies the data packet and writes it to the second buffer. For each data packet written, the hardware abstraction scheduling layer calculates the cyclic redundancy check (CRC) code of the data packet and stores it in the corresponding check fields of the two buffers. At predetermined time intervals, the hardware abstraction scheduling layer compares the check codes of the data packets at corresponding positions in the two buffers. If an inconsistency is found, the differing data packet is copied from the first buffer to overwrite the corresponding position in the second buffer, thereby maintaining the data consistency check between the two buffers.

[0046] Step S500: At the optimal handover execution time, perform an atomic data flow table migration operation to atomically switch the data flow exit of the current service network to the optimal target network while maintaining the continuity of upper-layer service connections.

[0047] Specifically, when the hardware timer reaches the optimal switching execution time of the hardware pre-preparation instruction set set in step S400, it generates an interrupt signal, which is captured by the atomic switching execution module. In this interrupt service routine, the atomic switching execution module first disables all interrupts of the local CPU to ensure that subsequent operations cannot be interrupted by other processes or interrupts. Then, the atomic switching execution module reads the transport layer protocol type, source port number, destination port number, and the network layer address of the current data flow from the service requirement feature vector output in step S130, using these fields as lookup keys to locate the corresponding entry in the data flow forwarding table. This entry stores the current output interface identifier. The atomic switching execution module modifies the value of the output interface identifier field in this entry from the interface identifier corresponding to the current serving network to the interface identifier corresponding to the optimal target network. This modification is completed by executing a single compare and swap assembly instruction, ensuring that the modification process cannot be interrupted within a CPU instruction cycle. After the modification is completed, the atomic switching execution module re-enables the interrupts of the local CPU. Since the network layer address, transport layer port, and encryption context of the data stream remain unchanged, the state of the upper-layer service connection remains unchanged, and the application layer continues to send and receive data through the new output interface without being aware of it. After the atomic handover is completed, the hardware abstraction scheduling layer immediately stops the double-buffered rolling storage mechanism, releases the memory space occupied by the first buffer, and associates the second buffer with the sending queue of the optimal target network. Data is then sent from the current read pointer position in the second buffer, achieving zero loss of service data during the handover process.

[0048] In one possible implementation, the data flow exit atom of the current service network is switched to the optimal target network while maintaining the continuity of the upper-layer service connection. Step S500 further includes step S510, in the data flow forwarding table, replacing the output interface identifier atom associated with the current data flow with the output interface identifier of the optimal target network, while keeping the network layer address, transport layer port and encryption context of the data flow unchanged. The replacement operation is completed within an uninterrupted instruction cycle.

[0049] Specifically, after the atomic handover execution module locates the target entry in the data flow forwarding table, the data structure of this entry includes a field for storing the output interface identifier. The interface identifier value corresponding to the current serving network is a first value, and the interface identifier value corresponding to the optimal target network is a second value different from the first value. The atomic handover execution module first clears the interrupt enable bit in the current processor status register, disabling all maskable interrupts. Subsequently, the module directly writes the interface identifier value corresponding to the optimal target network, i.e., the second value, into the output interface identifier field of the target entry. The underlying assembly instruction corresponding to this write operation is a register-to-memory instruction. Since the processor does not respond to any interrupts during the execution of this instruction, the replacement operation is atomic. After the write is completed, the atomic handover execution module immediately resets the interrupt enable bit, restoring interrupt response. Throughout the entire replacement operation, other fields in the data flow forwarding table, including the source Internet Protocol address, destination Internet Protocol address, source port number, destination port number, protocol type, and pointer to the encryption context of the data flow, remain unchanged, thereby ensuring the continuity of upper-layer service connections.

[0050] This application embodiment solves the technical problem of insufficient timeliness and reliability of network handover control in existing intelligent converged terminals by real-time collection of link status, terminal resource status, and service quality requirements of current services. It uses an online incremental learning prediction model based on time series to predict the service quality feasible domain of each candidate network within a future time window and calculates the feasible domain margin of the current service network. When the margin is lower than a dynamically set safety threshold, an active handover is triggered. After responding to the trigger signal, the optimal target network and handover execution time are determined by comprehensively considering link status, resource status, service requirements, and prediction results. Before the handover time arrives, a pre-preparation instruction is issued to the hardware abstract scheduling layer. At the handover time, the data flow table is migrated atomically to seamlessly switch the data flow exit to the target network and maintain the continuity of upper-layer service connections. These technical means solve the technical problem of insufficient timeliness and reliability of handover decision-making in the network handover control of existing intelligent converged terminals, and achieve the technical effect of improving the timeliness and reliability of handover decision-making.

[0051] In the above text, refer to Figure 1 A method for switching control of intelligent converged terminals supporting multiple network access according to an embodiment of the present invention is described in detail. Next, reference will be made to... Figure 2 This invention describes a smart converged terminal switching control system that supports multiple network access, according to an embodiment of the present invention.

[0052] The intelligent converged terminal handover control system supporting multiple network access according to embodiments of the present invention addresses the technical problem of insufficient timeliness and reliability of handover decisions in existing intelligent converged terminal network handover control, thereby improving the timeliness and reliability of handover decisions. The intelligent converged terminal handover control system supporting multiple network access includes: a status parameter acquisition module 10, an active handover trigger signal generation module 20, a decision module 30, a hardware pre-preparation instruction set issuance module 40, and a network handover module 50.

[0053] The status parameter acquisition module 10 is used to collect in real time the link status parameter sets of at least two heterogeneous networks accessed by the intelligent converged terminal, the resource status parameter set of the terminal itself, and the service demand feature vector of the currently carried service; the active handover trigger signal generation module 20 is used to input the link status parameter sets into an online incremental learning prediction model based on time series, predict the end-to-end service quality feasible region of each candidate network within a future time window, and determine whether the service quality feasible region margin of the current service network is lower than a preset safety threshold based on the service demand feature vector. If it is lower, an active handover trigger signal is generated; the decision module 30 is used to respond to the active handover trigger signal. The switching trigger signal, based on the link status parameters, resource status parameters, service demand feature vector, and service quality feasible domain prediction results, determines the optimal target network identifier and the optimal switching execution time; the hardware pre-preparation instruction set issuing module 40 is used to issue a hardware pre-preparation instruction set to the hardware abstraction scheduling layer of the intelligent converged terminal before the optimal switching execution time arrives, according to the optimal target network identifier; the network switching module 50 is used to perform an atomic-level data flow table migration operation at the optimal switching execution time, atomically switching the data flow exit of the current service network to the optimal target network, while maintaining the continuity of upper-layer service connections.

[0054] The detailed description of the specific configuration of the hardware pre-preparation instruction set distribution module 40 is explained as follows: As mentioned above, the hardware pre-preparation instruction set distribution module 40 may further include: the hardware pre-preparation instruction set includes: wake-up and parameter pre-configuration of the target communication module, pre-establishment of the wireless bearer, and pre-allocation of the data buffer; the hardware pre-preparation instruction set also includes an instruction to start a double-buffer rolling storage mechanism, wherein the data to be sent by the current service network is simultaneously written into the buffer pre-allocated to the target network, and the data consistency verification of the two buffers is maintained. After the atomic handover is completed, the data is immediately resumed from the buffer of the target network to achieve zero loss of service data during the handover process.

[0055] The specific configuration of the state parameter acquisition module 10 is described in detail below: As mentioned above, the state parameter acquisition module 10 may further include: a service requirement feature vector acquisition unit for capturing the transport layer protocol type, port number, data packet arrival interval, and packet length distribution of the current data stream as raw feature parameters; inputting the raw feature parameters into a preset service classifier and outputting a service type label; querying a preset service quality constraint mapping table based on the service type label to obtain the maximum latency tolerance, minimum bandwidth requirement, maximum packet loss tolerance, and service priority, which together constitute the service requirement feature vector.

[0056] The specific configuration of the active handover trigger signal generation module 20 is described in detail below: As mentioned above, the link state parameter set is input into an online incremental learning prediction model based on time series to predict the end-to-end service quality feasible region of each candidate network within a future time window. The active handover trigger signal generation module 20 may further include: a prediction unit for inputting the link state parameter set into the online incremental learning prediction model based on time series, and for adding a confidence interval to each predicted value based on the statistical distribution of historical prediction errors, wherein the upper and lower bounds of the confidence interval constitute the feasible region boundary of the corresponding network in the service quality index dimension; and a determination unit for determining that the corresponding network has service quality feasibility within the current time window when the service demand constraints within the future time window completely fall within the feasible region boundary.

[0057] The parameters of the online incremental learning prediction model are updated online based on historical prediction errors. The prediction unit may further include: a prediction error calculation subunit for obtaining actual measured values ​​of link status, comparing the actual measured values ​​with the model prediction values, and calculating the prediction error; and a model parameter correction subunit for using the prediction error to correct the model parameters through a recursive optimization algorithm, so that the model's response to changes in network status lags by no more than one prediction period.

[0058] The detailed description of the specific configuration of the decision module 30 is explained as follows: As mentioned above, in response to the active handover trigger signal, based on the link state parameters, resource state parameters, service demand feature vector, and service quality feasible domain prediction results, the decision module 30 determines the optimal target network identifier and the optimal handover execution time. The decision module 30 may further include: a decision unit that uses the link state parameters, resource state parameters, service demand feature vector, and service quality feasible domain prediction results to jointly construct a state space, uses the handover comprehensive benefit function as a reward signal, calls a decision engine based on deep reinforcement learning, and outputs the optimal target network identifier and the optimal handover execution time; wherein, the handover comprehensive benefit function integrates bandwidth gain, latency improvement, power consumption reduction, handover overhead, and ping-pong handover inhibition factor, and the weight of each factor is dynamically adjusted according to the service demand feature vector.

[0059] The decision-making unit, which determines the optimal target network identifier and the optimal switching execution time, may further include: a search time axis construction subunit, which constructs a search time axis with the current time as the starting point and a future preset time window as the ending point, using a discretized time step; a feasible boundary space construction subunit, which compares the service quality feasible domain prediction results with the business demand feature vector for each candidate network, and selects network-time pairs that satisfy the business constraints at each time point on the time axis, thus forming a feasible boundary space; and a policy network inference subunit, which performs policy network inference within the feasible boundary space, with the switching comprehensive benefit function as the objective, to solve for the network identifier that maximizes the benefit and the corresponding switching execution time; wherein, when the feasible boundary space is empty, the decision engine outputs the current service network as the target network, outputs a warning flag, and triggers an emergency handling process.

[0060] The strategy network inference subunit may further include: the switching comprehensive benefit function considers the expected dwell time after switching during the search process, wherein if a candidate network is expected to maintain a service quality feasible domain for more than a preset long time threshold after switching in, an additional long-term stability reward is added to the benefit function; if it is predicted that another switch will be needed in a short period of time after switching in, a discount penalty for future switching costs is added.

[0061] The network switching module 50 is described in detail below: As mentioned above, the network switching module 50 atomically switches the data flow exit of the current service network to the optimal target network while maintaining the continuity of the upper-layer service connection. The network switching module 50 may further include: a replacement unit used to atomically replace the output interface identifier associated with the current data flow in the data flow forwarding table with the output interface identifier of the optimal target network, while keeping the network layer address, transport layer port and encryption context of the data flow unchanged. The replacement operation is completed within an uninterrupted instruction cycle.

[0062] The intelligent converged terminal switching control system supporting multiple network access provided in the embodiments of the present invention can execute the intelligent converged terminal switching control method supporting multiple network access provided in any embodiment of the present invention, and has the corresponding functional modules and beneficial effects of the method execution.

[0063] Although this application makes various references to certain modules in the system according to the embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The various units and modules included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of each functional unit are only for easy distinction between each other and are not used to limit the scope of protection of this invention.

[0064] The specific embodiments described above do not constitute a limitation on the scope of protection of this application. Those skilled in the art should understand that various modifications, combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application. In some cases, the actions or steps described in this application can be performed in a different order than that shown in the embodiments and still achieve the desired results. Furthermore, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

Claims

1. A method for switching control of intelligent converged terminals supporting multiple network accesses, characterized in that, include: Real-time collection of link status parameter sets of at least two heterogeneous networks accessed by the intelligent converged terminal, resource status parameter sets of the terminal itself, and service requirement feature vectors of the services currently carried. The link state parameter set is input into the online incremental learning prediction model based on time series to predict the end-to-end service quality feasible region of each candidate network within the future time window. Based on the service demand feature vector, it is determined whether the service quality feasible region margin of the current service network is lower than the preset safety threshold. If it is lower, an active handover trigger signal is generated. In response to the active handover trigger signal, based on the link status parameters, resource status parameters, service demand feature vector, and service quality feasible domain prediction results, the optimal target network identifier and optimal handover execution time are determined. Before the optimal handover execution time arrives, a hardware pre-preparation instruction set is issued to the hardware abstraction scheduling layer of the intelligent fusion terminal according to the optimal target network identifier; At the optimal handover execution time, an atomic data flow table migration operation is performed to atomically switch the data flow exit of the current service network to the optimal target network while maintaining the continuity of upper-layer service connections; Specifically, in response to the active handover trigger signal, based on the link status parameters, resource status parameters, service demand feature vector, and service quality feasible region prediction results, the optimal target network identifier and optimal handover execution time are determined, including: The state space is composed of the link state parameters, resource state parameters, service demand feature vector, and service quality feasible region prediction results. The switching comprehensive benefit function is used as the reward signal. The decision engine based on deep reinforcement learning is invoked to output the optimal target network identifier and the optimal switching execution time. The handover comprehensive benefit function integrates bandwidth gain, latency improvement, power consumption reduction, handover overhead, and ping-pong handover suppression factor, and the weight of each factor is dynamically adjusted according to the service demand feature vector. Among them, the optimal target network identifier and the optimal handover execution time are determined as follows: Starting from the current moment and ending at a future preset time window, a search time axis is constructed using discrete time steps; For each candidate network, the service quality feasible region prediction results are compared with the business demand feature vector, and the network-time pairs that satisfy the business constraints at each time point on the time axis are selected to form the feasible boundary space. Within the feasible boundary space, with the switching comprehensive benefit function as the objective, policy network reasoning is performed to solve for the network identifier that maximizes the benefit and the corresponding switching execution time. When the feasible boundary space is empty, the decision engine outputs the current service network as the target network, outputs an early warning flag, and triggers an emergency handling process.

2. The intelligent converged terminal switching control method supporting multiple network access according to claim 1, characterized in that, The hardware pre-preparation instruction set includes: wake-up and parameter pre-configuration of the target communication module, pre-establishment of the wireless bearer, and pre-allocation of the data buffer. The hardware pre-preparation instruction set also includes an instruction to start a dual-buffered rolling storage mechanism, in which the data to be sent from the current service network is simultaneously written into the buffer pre-allocated to the target network, and the data consistency verification between the two buffers is maintained. After the atomic handover is completed, the data is immediately resumed from the buffer of the target network, achieving zero loss of service data during the handover process.

3. The intelligent converged terminal switching control method supporting multiple network access according to claim 1, characterized in that, The method for obtaining the service requirement feature vector of the currently carried service is as follows: Capture the transport layer protocol type, port number, packet arrival interval, and packet length distribution of the current data stream as raw feature parameters; The original feature parameters are input into a preset business classifier, which outputs a business type label. The service type label is used to query the preset service quality constraint mapping table to obtain the maximum latency tolerance, minimum bandwidth requirement, maximum packet loss tolerance, and service priority, which together constitute the service requirement feature vector.

4. The intelligent converged terminal switching control method supporting multiple network access according to claim 1, characterized in that, The link state parameter set is input into a time-series-based online incremental learning prediction model to predict the end-to-end service quality feasible region for each candidate network within a future time window, including: The link state parameter set is input into the online incremental learning prediction model based on time series. Based on the statistical distribution of historical prediction errors, a confidence interval is added to each prediction value. The upper and lower bounds of the confidence interval constitute the feasible domain boundary of the corresponding network in the dimension of service quality indicators. When the business demand constraints within a future time window fall completely within the feasible domain boundary, the corresponding network is determined to have service quality feasibility within this time window.

5. The intelligent converged terminal switching control method supporting multiple network access according to claim 4, characterized in that, The parameters of the online incremental learning prediction model are updated online based on historical prediction errors, including: Obtain the actual measured values ​​of the link status, compare the actual measured values ​​with the model prediction values, and calculate the prediction error; The model parameters are corrected using a recursive optimization algorithm based on the prediction error, so that the model's response to changes in network state lags by no more than one prediction period.

6. The intelligent converged terminal switching control method supporting multiple network access according to claim 1, characterized in that, The switching comprehensive benefit function considers the expected dwell time after switching during the search process. If a candidate network is expected to maintain a service quality feasible domain for more than a preset long time threshold after switching in, an additional long-term stability reward is added to the benefit function; if it is predicted that another switch will be needed in a short period of time after switching in, a discount penalty for future switching costs is added.

7. The intelligent converged terminal switching control method supporting multiple network access according to claim 1, characterized in that, Atomically switch the data flow exit of the current service network to the optimal target network while maintaining the continuity of upper-layer service connections, including: In the data flow forwarding table, the output interface identifier associated with the current data flow is atomically replaced with the output interface identifier of the optimal target network, while keeping the network layer address, transport layer port and encryption context of the data flow unchanged. The replacement operation is completed within an uninterrupted instruction cycle.

8. A smart converged terminal switching control system supporting multiple network accesses, characterized in that, The system is used to implement the intelligent converged terminal handover control method supporting multiple network access as described in any one of claims 1-7, and the system includes: The status parameter acquisition module is used to collect in real time the link status parameter sets of at least two heterogeneous networks accessed by the intelligent converged terminal, the resource status parameter set of the terminal itself, and the service requirement feature vector of the currently carried service. The active handover trigger signal generation module is used to input the link state parameter set into the online incremental learning prediction model based on time series, predict the end-to-end service quality feasible domain of each candidate network within the future time window, and determine whether the service quality feasible domain margin of the current service network is lower than the preset safety threshold based on the service demand feature vector. If it is lower, an active handover trigger signal is generated. The decision module is used to respond to the active handover trigger signal and, based on the link status parameters, resource status parameters, service demand feature vector, and service quality feasible domain prediction results, decide the optimal target network identifier and the optimal handover execution time. The hardware pre-preparation instruction set issuing module is used to issue a hardware pre-preparation instruction set to the hardware abstract scheduling layer of the intelligent fusion terminal according to the optimal target network identifier before the optimal handover execution time arrives. The network switching module is used to perform an atomic data flow table migration operation at the optimal switching execution time, atomically switching the data flow exit of the current service network to the optimal target network, while maintaining the continuity of upper-layer service connections.

Images