Three-network integration intelligent switching system and method based on lightweight neural network

Abstract

The invention relates to the technical field of communication data processing, in particular to a three-network integration intelligent switching system and method based on a lightweight neural network, and the system comprises a data collection module, a prediction module, a credibility correction module, a decision control module and a firmware execution module. According to the method, the lightweight neural network model is deployed on the terminal side, and the future network quality is predicted, so that the switching decision is converted into prospective judgment from passive response; and stability evaluation and self-adaptive correction are carried out on a prediction result through a credibility correction mechanism to suppress abnormal fluctuation interference, so that an accurate switching strategy is generated based on a corrected decision index, and finally efficient execution is carried out through a firmware layer. The problems that network deterioration cannot be avoided prospectively, the switching decision response speed is low and the reliability is poor due to dependence on lagging judgment on the current state and insufficient model calculation efficiency are effectively solved.

Description

Technical Field

[0001] This invention relates to the field of communication data processing technology, and in particular to a three-network convergence intelligent handover system and method based on a lightweight neural network. Background Technology

[0002] With the deep integration and widespread deployment of heterogeneous access technologies such as high-speed cellular networks, next-generation wireless LANs, and various dedicated wireless networks, user terminals are facing a complex communication environment characterized by the coexistence of multiple network standards, intertwined and dynamically changing signal coverage. Simultaneously, ultra-high-definition video streams, real-time interactive applications, and mission-critical services place near-stringent demands on network latency, reliability, and continuity. Against this backdrop, traditional network switching mechanisms relying on fixed thresholds and delayed responses are no longer sufficient to meet these demands due to their inability to predict quality trends and their tendency to cause service interruptions and resource waste. Meanwhile, existing intelligent solutions based on cloud computing or complex neural networks are limited by terminal computing power, model complexity, and system integration overhead, making it difficult to achieve efficient and real-time decision-making and execution at the actual terminal side. These factors collectively constitute the core challenges in improving the seamless experience and reliable service of heterogeneous networks.

[0003] Chinese Patent Publication No. CN113630830A discloses a network switching method, apparatus, and device based on a multi-attribute fuzzy neural network. The method includes: acquiring a set of candidate switching access points for a terminal device; if the set of candidate switching access points for the terminal device includes multiple candidate switching access points, then inputting the motion attribute information of the terminal device and the network attribute information of each candidate switching access point into a trained first target fuzzy neural network to obtain the switching probability of each candidate switching access point, and determining the candidate switching access point with the highest switching probability as the target switching access point of the terminal device. The first target fuzzy neural network is trained based on multiple sets of first historical switching sample data that have been updated in real time. Each set of first historical switching sample data includes the input motion attribute information and network attribute information, as well as the corresponding output switching probability of each candidate switching access point. The switching probability is used to characterize the probability of the terminal device switching to a candidate switching access point; if the set of candidate switching access points for the terminal device includes one candidate switching access point, then determining that candidate switching access point as the target switching access point of the terminal device; determining whether the target switching access point meets the network switching conditions; and when the target switching access point meets the network switching conditions, switching the terminal device from the current access point to the target switching access point.

[0004] Therefore, the existing technology has the following problems: the fuzzy neural network it uses is essentially based on historical data to classify and make decisions about the current network state, lacking the ability to proactively predict future network quality changes, and thus cannot achieve forward-looking operations to avoid the risk of business interruption; its decision-making is highly dependent on modeling a specific combination of motion attributes and network attributes, and when the terminal motion pattern or network environment characteristics exceed the range of historical training data, the accuracy and reliability of the model's decision-making are prone to significant decline; the fuzzy neural network it uses may generate high computational latency and power consumption on resource-constrained terminal devices, making it difficult to support millisecond-level real-time switching decisions and execution. Summary of the Invention

[0005] To address this, the present invention provides a three-network convergence intelligent handover system and method based on lightweight neural networks. This system uses a lightweight neural network model deployed on the terminal side to predict future network quality and introduces a reliability correction mechanism based on prediction stability to overcome the problems in the prior art, such as the inability to proactively avoid network deterioration, low handover decision response speed, and poor reliability due to reliance on delayed judgment of the current state and insufficient model computation efficiency.

[0006] To achieve the above objectives, on the one hand, the present invention provides a three-network convergence intelligent handover system based on a lightweight neural network, comprising: The data acquisition module is used to collect network quality parameters of cellular networks, wireless LANs and private networks in real time on the terminal side. The prediction module, which is a hardware acceleration unit deployed on the terminal side, is used to construct a feature input of historical time series based on the network quality parameters, and input it into a preset lightweight neural network model to predict the network quality of the cellular network, the wireless local area network and the dedicated network in the future time window, and output the corresponding predicted quality index. The credibility correction module is used to generate a credibility quantification factor to reflect the prediction stability based on the change characteristics of the prediction quality index over several consecutive prediction periods, and to perform joint correction or suppression correction on the prediction quality index based on the change trend determination result of the credibility quantification factor within the future time window, so as to obtain a decision index for switching determination. The decision control module is used to generate network switching strategies based on the numerical range distribution and relative ranking relationship of the decision index, combined with the current service. The firmware execution module is used to generate corresponding protocol stack control instructions based on the network switching strategy, and execute the protocol stack control instructions through the hardware acceleration unit to complete the switching between the cellular network, the wireless local area network and the private network.

[0007] Furthermore, the data acquisition module includes: The parameter acquisition unit is used to acquire signal strength, round-trip time, available bandwidth and packet loss rate from the driver layer of the cellular network, the wireless local area network and the dedicated network, respectively. The time alignment unit is used to perform timestamp alignment and sampling period unification processing on the signal strength, round-trip delay, available bandwidth and packet loss rate from different network sources to form the network quality parameters.

[0008] Furthermore, the prediction module includes: The feature construction unit is used to construct the network quality parameters into a feature matrix containing multiple historical time points based on a preset time sliding window; The model calculation unit is used to call the lightweight neural network model, calculate the feature matrix, and output the predicted quality index corresponding to the cellular network, wireless local area network and dedicated network in the future time window.

[0009] Furthermore, the credibility correction module includes: A fluctuation analysis unit is used to calculate the fluctuation amplitude and dispersion of the prediction quality index within a consecutive preset number of prediction periods; A credibility calculation unit is used to generate the corresponding credibility quantification factor based on the fluctuation amplitude and dispersion. A correction control unit is used to jointly correct the prediction quality index when the confidence quantification factor meets a preset stability condition, or to suppress the prediction quality index when the stability condition is not met, so as to obtain the decision index.

[0010] Furthermore, the decision control module includes: An interval determination unit is used to determine whether the cellular network, the wireless local area network, or the dedicated network meets the handover triggering conditions based on the decision index falling within a preset decision interval, and to generate a corresponding candidate network set. The sorting and comparison unit is used to sort the corresponding decision indices in the candidate network set from largest to smallest, and mark the first decision index as the primary index and the second decision index as the alternative index. The strategy selection unit is used to combine the current service type and the difference between the primary index and the alternative index to generate a single-network switching strategy, a multi-network concurrent strategy, or a pre-switching strategy to obtain the network switching strategy.

[0011] Furthermore, the strategy selection unit includes: A single-network handover generation subunit is used to generate the single-network handover strategy when the primary selection index is higher than a preset first handover threshold and the difference between the primary selection index and the alternative index is greater than a preset difference threshold. A multi-network concurrent generation subunit is used to generate the multi-network concurrent strategy when both the primary index and the alternative index are higher than a preset second concurrency threshold, and the difference is less than or equal to the preset difference threshold. The pre-switching generation subunit is used to generate the pre-switching strategy when the primary selection index is higher than the preset third pre-switching threshold and the latency sensitivity level of the current service meets the preset pre-switching conditions.

[0012] Furthermore, the strategy selection unit also includes: The indicator extraction subunit is used to extract the maximum allowable latency of a service from the service data packets at the terminal side. The level calculation subunit is used to calculate the corresponding latency sensitivity level based on the maximum allowable latency and according to the preset level classification rules.

[0013] Furthermore, the firmware execution module includes: The policy parsing unit is used to parse the network handover policy to determine the target network, handover type, and handover execution priority; The instruction generation unit is used to generate a sequence of hardware control instructions that matches the network communication protocol stack of the target network based on the target network, the switching type and the priority, so as to obtain the protocol stack control instructions. The driver execution unit is used to call the dedicated driver interface of the hardware acceleration unit to execute the protocol stack control instructions, so as to control the corresponding radio frequency front-end and baseband processing unit on the terminal side to complete the switching operation to the target network.

[0014] Furthermore, the instruction generation unit includes: The protocol matching subunit is used to determine the corresponding network communication protocol stack type according to the network type of the target network, and select a control instruction template that matches the network communication protocol stack type from a preset protocol instruction template set; The parameter filling subunit is used to fill the configurable parameters in the control instruction template based on the priority of the switching execution and the current network operating status of the terminal, so as to generate specific protocol stack control instructions. The timing arrangement subunit is used to sort and combine the protocol stack control instructions according to the dependencies between them, so as to form the protocol stack control instructions that can be executed sequentially.

[0015] On the other hand, the present invention also provides a method for intelligent handover of three networks based on lightweight neural networks, comprising: Real-time collection of network quality parameters for cellular networks, wireless LANs, and dedicated networks at the terminal side; Based on the network quality parameters, a feature input containing historical time series is constructed and input into a preset lightweight neural network model to predict the network quality of the cellular network, the wireless local area network and the dedicated network in the future time window, and output the corresponding predicted quality index. Based on the change characteristics of the prediction quality index over several consecutive prediction periods, a reliable quantification factor is generated to reflect the prediction stability. Based on the change trend determination result of the reliable quantification factor within the future time window, the prediction quality index is jointly corrected or suppressed to obtain a decision index for switching determination. Based on the numerical range distribution and relative ranking relationship of the decision index, and combined with the current service, a network switching strategy is generated. Based on the network switching strategy, corresponding protocol stack control instructions are generated and executed by the hardware acceleration unit to complete the switching between the cellular network, the wireless local area network, and the private network.

[0016] Compared with existing technologies, the advantages of this invention are as follows: by continuously collecting key quality parameters of multi-standard networks at the terminal side and constructing feature inputs in a time-series manner, the short-term fluctuations and medium-term trends of network states can be uniformly characterized; based on lightweight neural networks, the quality of different networks within future time windows is predicted, transforming the switching decision from a passive response to a forward-looking judgment; furthermore, by quantifying the fluctuation amplitude and stability of the predicted quality index within a continuous prediction period, a credibility correction mechanism is formed, which suppresses the prediction results in unstable or abrupt scenarios and enhances them in stable evolution scenarios, thereby avoiding misleading switching decisions due to single prediction anomalies; on this basis, the decision control module performs interval judgment and ranking comparison on the corrected decision index, enabling network switching behavior to simultaneously consider the absolute quality level and the relative advantages among multiple networks; finally, the firmware execution module maps high-level policies into control instructions that match specific protocol stacks and hardware states, enabling the decision results to be executed in a controllable timing and priority manner. While ensuring timely handover, it effectively reduces the probability of frequent and erroneous handovers, improves connection stability and service continuity in multi-network coexistence environments, and effectively solves the problems of being unable to proactively avoid network deterioration, low handover decision response speed, and poor reliability due to reliance on delayed judgment of the current state and insufficient model calculation efficiency.

[0017] Furthermore, by performing high-precision time synchronization and sampling alignment on the multi-dimensional native parameters from the heterogeneous network driver layer, key indicators such as signal strength, round-trip delay, available bandwidth, and packet loss rate are ensured to be comparable and consistent in timing under a strictly unified spatiotemporal benchmark. This enables these parameters, which have an inherent coupling relationship between the physical layer and the data link layer, to be analyzed and processed collaboratively, providing a reliable and consistent data foundation for subsequent neural network models to mine the causal relationship of network state evolution from historical time series and make high-quality predictions.

[0018] Furthermore, by organizing multi-dimensional network quality parameters with strict time alignment into a feature matrix in chronological order, the symbiotic and transmission relationships of physical layer and transport layer phenomena such as signal strength fading, round-trip delay accumulation, bandwidth fluctuations, and packet loss events on a continuous time scale can be structurally characterized. The lightweight neural network model, through deep nonlinear transformation and attention weighting of this spatiotemporal feature matrix, can effectively learn and extract the dynamic coupling patterns of different network parameters in the short-term evolution process, which are determined by wireless channel characteristics and protocol stack behavior. This enables highly timely and accurate prediction of the future comprehensive state of the network, providing a key and reliable input basis for subsequent intelligent decision-making.

[0019] Furthermore, by quantitatively analyzing the fluctuation amplitude and dispersion of the prediction quality index over continuous time, the system can objectively assess the inherent uncertainty of the neural network prediction results. Based on this assessment, the generated credible quantification factor enables the system to distinguish between high-confidence steady-state predictions and transient fluctuations caused by random disturbances. Accordingly, it performs adaptive weighted smoothing or suppression correction on the original prediction index, transforming the time series statistical characteristics in the prediction model output into actionable decision information. This significantly enhances the reliability of the final decision index in representing the evolution trend of the real network state, effectively suppresses unnecessary switching oscillations caused by instantaneous model misjudgments or sudden channel noise, and improves the robustness and accuracy of the overall switching strategy.

[0020] Furthermore, by constructing a dynamically adjustable hierarchical decision-making framework, the system collaboratively matches and quantifies the modified decision index reflecting the overall network status with the latency tolerance determined by the inherent attributes of service data packets and the target differences of different handover actions (such as single-network handover, concurrent, or pre-handover). This framework initially screens the network quality status through preset decision intervals, then evaluates the relative advantages between networks based on index ranking and difference relationships, and finally precisely triggers the most suitable handover strategy by combining the transmission characteristics of the service itself. This enables a tight coupling between the quantitative assessment of network status, the objective constraints of service requirements, and the execution objectives of handover actions, thereby significantly improving the targeting, accuracy, and reliability of network handover strategies.

[0021] Furthermore, by transforming the network handover strategy generated by upper-layer intelligent decision-making into a sequence of hardware control instructions that strictly adheres to the target network communication protocol stack specifications and the timing dependencies of underlying hardware operations, a precise and efficient mapping from decision logic to physical layer operations is achieved. The system dynamically adjusts the execution time limits and resource scheduling in the instruction sequence based on the priority parameters in the strategy, and adapts specific radio frequency and baseband configuration parameters based on the real-time network status of the terminal. This ensures that the upper-layer strategy intent, the inherent constraints of the communication protocol, and the physical characteristics of the hardware units work in synergy, guaranteeing that for different modes such as pre-handover and hard handover, RF front-end reconfiguration, link reconstruction, and data path switching can be completed with optimal timing while meeting the protocol stack state machine transition requirements. Ultimately, this achieves millisecond-level low-latency stable execution of the entire intelligent handover system's decision-making closed loop.

[0022] Furthermore, by transforming the multidimensional heterogeneous network parameters, which reflect the physical and transport layer states of the wireless channel and are collected in real time at the edge, into structured features that characterize their historical temporal evolution, and using a lightweight neural network model to mine the inherent coupling and transmission relationships of these parameters in the continuous time dimension, the system predicts their future state. The system then performs quantitative evaluation and adaptive correction based on statistical stability on the prediction results. Finally, based on the corrected decision index and the inherent attributes of the service, it generates and executes handover instructions that strictly match the underlying communication protocol stack and hardware operations. This achieves an end-to-end closed loop from multi-source network data perception, temporal correlation feature extraction, uncertainty quantification processing, to precise strategy generation and execution, significantly improving the timeliness, accuracy, and service adaptability of network handover in complex mobile scenarios. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of a three-network convergence intelligent handover system based on a lightweight neural network, as described in this embodiment. Figure 2 This is a schematic diagram of the prediction module in this embodiment; Figure 3 This is a schematic diagram of the credibility correction module in this embodiment; Figure 4 This is a flowchart of the intelligent switching method for triple-play integration based on lightweight neural networks in this embodiment. Detailed Implementation

[0024] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0025] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0026] Please see Figure 1 As shown, this is a schematic diagram of a three-network convergence intelligent handover system based on a lightweight neural network, one aspect of this embodiment. This embodiment provides a three-network convergence intelligent handover system based on a lightweight neural network, including: Data acquisition module 1 is used to collect network quality parameters corresponding to cellular networks, wireless local area networks and private networks in real time on the terminal side; Prediction module 2, which is a hardware acceleration unit deployed on the terminal side and connected to the data acquisition module, is used to construct a feature input of historical time series based on the network quality parameters, and input it into a preset lightweight neural network model to predict the network quality of the cellular network, the wireless local area network and the dedicated network in the future time window, and output the corresponding predicted quality index. The credibility correction module 3 is connected to the prediction module and is used to generate a credibility quantification factor to reflect the prediction stability based on the change characteristics of the prediction quality index in a continuous number of prediction periods. Based on the change trend judgment result of the credibility quantification factor in the future time window, the prediction quality index is jointly corrected or suppressed to obtain a decision index for switching judgment. The decision control module 4, which is connected to the credibility correction module, is used to generate a network switching strategy based on the numerical range distribution and relative ranking relationship of the decision index and in combination with the current service. Firmware execution module 5, which is connected to decision control module, is used to generate corresponding protocol stack control instructions based on the network switching strategy, and execute the protocol stack control instructions through hardware acceleration unit to complete the switching between the cellular network, the wireless local area network and the private network.

[0027] In this embodiment, "terminal side" specifically refers to the integration of the entire set of logic for data acquisition, prediction, decision-making, and execution into the user device itself, such as a smartphone, vehicle communication unit, or industrial IoT gateway. The hardware acceleration unit refers to a dedicated computing chip built into the terminal, such as a neural network processing unit, digital signal processor, or graphics processor with corresponding computing power. This chip efficiently carries and runs the lightweight neural network model, enabling real-time, low-power forward inference calculations of network quality parameters. This allows the entire intelligent handover system to autonomously complete rapid and accurate network assessment and handover decisions without relying on cloud servers, especially when users experience changes in cellular network, wireless LAN, and dedicated network coverage and signal fluctuations during their movement.

[0028] By continuously collecting key quality parameters of multi-standard networks at the terminal side and constructing feature inputs in a time-series manner, the short-term fluctuations and medium-term trends of network states can be uniformly characterized. Based on a lightweight neural network, the quality of different networks within future time windows is predicted, transforming switching decisions from passive response to proactive judgment. Furthermore, by quantifying the fluctuation amplitude and stability of the predicted quality index within a continuous prediction period, a reliability correction mechanism is formed, suppressing prediction results in unstable or abrupt scenarios and enhancing them in stable evolution scenarios, thereby avoiding misleading switching decisions due to single prediction anomalies. On this basis, the decision control module performs interval judgment and ranking comparison on the corrected decision index, ensuring that network switching behavior simultaneously considers absolute quality levels and the relative advantages among multiple networks. Finally, the firmware execution module maps high-level policies into control instructions that match specific protocol stacks and hardware states, enabling the decision results to be executed in a controllable timing and priority manner. While ensuring timely handover, it effectively reduces the probability of frequent and erroneous handovers, improves connection stability and service continuity in multi-network coexistence environments, and effectively solves the problems of being unable to proactively avoid network deterioration, low handover decision response speed, and poor reliability due to reliance on delayed judgment of the current state and insufficient model calculation efficiency.

[0029] Specifically, the data acquisition module includes: The parameter acquisition unit is used to acquire signal strength, round-trip time, available bandwidth and packet loss rate from the driver layer of the cellular network, the wireless local area network and the dedicated network, respectively. A time alignment unit, connected to a parameter acquisition unit, is used to perform timestamp alignment and sampling period unification processing on the signal strength, round-trip delay, available bandwidth, and packet loss rate from different network sources to form the network quality parameters.

[0030] In this embodiment, the parameter acquisition unit periodically reads raw network status information from the physical layer or data link layer drivers of different networks through the standard interface or dedicated driver application programming interface provided by the terminal operating system kernel. Specifically, for cellular networks, it obtains the received signal reference power value and the round-trip time measured by layer 3 signaling through the radio resource control protocol stack interface; for wireless local area networks, it obtains the received signal strength indication, the available bandwidth estimated from the probe response or channel state information, and the packet loss rate statistically obtained by the transport layer protocol through the 802.11 media access control management interface; for dedicated networks, it obtains the field strength, signal-to-noise ratio, and link reliability indicators through the management plane of its specific communication protocol stack. The time alignment unit maintains a global time axis based on the system's high-precision clock. When asynchronous parameter data with its local timestamp arrives from each network, the unit first corrects each timestamp to the global time reference according to the clock synchronization mechanism. Then, it uses a fixed-interval resampling method, such as 20 milliseconds as a common sampling period, and uses a linear interpolation algorithm to interpolate and align all parameter sequences. Finally, it generates a multi-dimensional network quality parameter vector with strictly synchronized timestamps and completely consistent sampling times for use by the subsequent prediction module.

[0031] By performing high-precision time synchronization and sampling alignment on multi-dimensional native parameters from heterogeneous network driver layers, key indicators such as signal strength, round-trip time, available bandwidth, and packet loss rate are ensured to be comparable and consistent in timing under a strictly unified spatiotemporal benchmark. This enables these parameters, which have an inherent coupling relationship between the physical layer and the data link layer, to be analyzed and processed collaboratively, providing a reliable and consistent data foundation for subsequent neural network models to mine the causal relationship of network state evolution from historical time series and make high-quality predictions.

[0032] Please see Figure 2 As shown, this is a schematic diagram of the prediction module in this embodiment. In this embodiment, the prediction module 2 includes: Feature construction unit 21 is used to construct the network quality parameters into a feature matrix containing multiple historical time points based on a preset time sliding window; The model calculation unit 22, which is connected to the feature construction unit, is used to call the lightweight neural network model, calculate the feature matrix, and output the predicted quality index corresponding to the cellular network, wireless local area network and dedicated network in the future time window.

[0033] In this embodiment, the feature construction unit organizes and encodes network quality parameters according to a preset time sliding window. The window length of the time sliding window is N sampling periods, and the sliding step size is 1 sampling period. Within each sampling period, network quality parameters of cellular network, wireless LAN, and dedicated network are acquired respectively. The network quality parameters include received signal strength, round-trip time, available bandwidth, and packet loss rate.

[0034] The feature construction unit first normalizes the network quality parameters of different dimensions, and then concatenates the network quality parameters from N consecutive sampling periods in chronological order according to a preset parameter arrangement, constructing a feature matrix of dimension M×N, where M is the number of feature dimensions of the network quality parameters. M depends on the number of network types involved in the modeling and the types of quality parameters corresponding to each network, while N depends on the rate of change of network state and the prediction time domain requirements. Typically, M is set between 12 and 20, and N is set between 5 and 30; in this embodiment, M is set to 12 and N is set to 10 to ensure the integrity of the model input information while taking into account both computational complexity and real-time requirements.

[0035] In this embodiment, the feature matrix is ​​used as the input to the lightweight neural network model in tensor form. The lightweight neural network model is a dedicated network based on an improved ShuffleNetV2 architecture. It includes an input layer consisting of 1×1 convolutions, four cascaded improved ShuffleNetV2 basic units, and a fully connected output layer. Each improved ShuffleNetV2 basic unit integrates a lightweight temporal attention submodule. After structural pruning and parameter optimization, the number of parameters in the lightweight neural network model is less than 1MB.

[0036] Furthermore, the lightweight neural network model uses post-training static quantization to quantize model weights and activation values ​​to INT8 precision, adapting to the integer arithmetic instruction set of the terminal-side hardware acceleration unit, thereby achieving sub-millisecond inference performance. The model computation unit calls the quantized model deployed on the terminal-side hardware acceleration unit to perform forward inference computation on the feature matrix. Through cascaded feature extraction and nonlinear mapping processes, it outputs the predicted quality index of cellular networks, wireless LANs, and dedicated networks within a future time window.

[0037] The prediction quality index is a scalar or vector used to characterize the overall communication quality of the network, and is used for subsequent reliability correction and network switching decision processing. The improved ShuffleNetV2 basic unit is based on the standard ShuffleNetV2 basic unit, with a lightweight temporal attention submodule embedded in its residual connection path. The temporal attention submodule aggregates temporal features through a one-dimensional global average pooling layer, and then generates a channel attention weight vector through a structure containing a dimensionality reduction fully connected layer, an activation function, and an up-dimensional fully connected layer. Finally, the weight vector is multiplied by the original feature map.

[0038] The preset time sliding window refers to a fixed-length time interval used for time-series aggregation of continuous network quality sampling data. Its length depends on the network state change rate, service latency sensitivity, and prediction stability requirements, and is typically set between 0.5 and 5 seconds. In this embodiment, it is set to 1 second, which effectively captures short-term network fluctuations while avoiding the introduction of excessive historical redundancy. The future time window's length is determined based on service latency sensitivity and the lead time requirement for handover decisions, and is typically set between 50 and 500 milliseconds. In this embodiment, it is set to 200 milliseconds, providing a suitable lead time for network quality prediction. This ensures sufficient timeliness of the prediction results to support proactive decisions such as pre-handover, while avoiding excessive uncertainty introduced by an excessively long prediction time domain, thus achieving an effective balance between prediction foresight and accuracy.

[0039] By organizing multi-dimensional network quality parameters with strict time alignment into a feature matrix in chronological order, the symbiotic and transmission relationships of physical layer and transport layer phenomena such as signal strength fading, round-trip delay accumulation, bandwidth fluctuations, and packet loss events on a continuous time scale can be structurally characterized. The lightweight neural network model, through deep nonlinear transformation and attention weighting of this spatiotemporal feature matrix, can effectively learn and extract the dynamic coupling patterns of different network parameters in the short-term evolution process, which are determined by wireless channel characteristics and protocol stack behavior. This enables highly timely and accurate prediction of the future comprehensive state of the network, providing a key and reliable input basis for subsequent intelligent decision-making.

[0040] Please see Figure 3 As shown, this is a schematic diagram of the credibility correction module in this embodiment. In this embodiment, the credibility correction module 3 includes: The fluctuation analysis unit 31 is used to calculate the fluctuation amplitude and dispersion of the prediction quality index within a consecutive preset number of prediction periods; The credibility calculation unit 32 is connected to the fluctuation analysis unit to generate the corresponding credibility quantification factor based on the fluctuation amplitude and dispersion. The correction control unit 33, which is connected to the credibility calculation unit, is used to jointly correct the prediction quality index when the credibility quantification factor meets the preset stability condition, or to suppress the prediction quality index when the stability condition is not met, so as to obtain the decision index.

[0041] In this embodiment, the fluctuation analysis unit uses the time series of the predicted quality index over a predetermined number of consecutive prediction periods as the analysis object. The predetermined number of analysis periods characterizes the stability of the prediction results over a short timescale. The fluctuation analysis unit calculates the difference between the maximum and minimum values ​​of the predicted quality index over the predetermined number of prediction periods to obtain the corresponding fluctuation amplitude, and further calculates the standard deviation of the predicted quality index relative to its mean to obtain the corresponding dispersion. The value of the predetermined number of analysis periods depends on the length of the prediction time window and the switching response speed requirements, and is typically set between 3 and 10. In this embodiment, it is set to 5, which can balance the smoothness of the prediction results and the timeliness of the switching response while avoiding excessive influence of a single abnormal prediction on the stability determination.

[0042] The credibility calculation unit generates a credibility quantification factor based on volatility and dispersion according to a preset weighted fusion rule. The weighted fusion rule involves normalizing volatility and dispersion separately, then constructing a comprehensive stability evaluation value using a linear weighting method. This comprehensive stability evaluation value is then mapped to a preset credibility interval to form a numerical credibility quantification factor, which characterizes the credibility of the prediction quality index within a continuous prediction period. The preset amplitude weight corresponding to the fluctuation amplitude depends on the transient fluctuation characteristics of the network quality parameters and the sensitivity of the service to sudden performance changes. It is usually set between 0.4 and 0.7. In this embodiment, it is set to 0.6, which can enhance the ability to suppress drastic fluctuations in the prediction results. The preset dispersion weight corresponding to the dispersion depends on the overall stability requirements of the prediction results in the time dimension. It is usually set between 0.3 and 0.6. In this embodiment, it is set to 0.4, which can reflect the overall consistency characteristics of the prediction quality index in a continuous period. The preset confidence interval refers to the standardized value range used to numerically map the comprehensive stability evaluation value. It is used to uniformly characterize the confidence level of the prediction quality index in a continuous prediction period. It is set in the form of a closed interval [0, 1]. The closer the value is to 1, the higher the stability of the prediction result in the time dimension. The closer the value is to 0, the greater the volatility of the prediction result. The comprehensive stability evaluation value is projected into the preset confidence interval through linear mapping or normalized mapping. During the mapping process, the relative magnitude relationship between different prediction results remains unchanged, so that the confidence quantification factor can be directly used for subsequent stability threshold comparison and correction strategy determination.

[0043] The correction control unit performs differentiated correction strategies on the predicted quality index based on the comparison between the reliability quantification factor and the preset stability threshold. When the reliability quantification factor is higher than the preset stability threshold, the correction control unit performs joint correction processing on the predicted quality index of the corresponding network, that is, it performs weighted smoothing fusion of the current predicted quality index and the historical predicted quality index to enhance the continuity of the prediction results. When the reliability quantification factor is lower than the stability threshold, the correction control unit performs suppression correction processing on the corresponding predicted quality index, such as reducing its weight in subsequent handover decisions or introducing a decay coefficient, to avoid frequent or false network handovers caused by abnormal prediction fluctuations. Among them, the preset current index weight corresponding to the current predicted quality index depends on the degree of dependence of the service on real-time network state changes, and is usually set between 0.5 and 0.8. In this embodiment, it is set to 0.7, which can ensure that the handover decision remains sufficiently sensitive to the latest network state changes. The preset historical index weight corresponding to the historical predicted quality index depends on the continuity characteristics of network state evolution, and is usually set between 0.2 and 0.5. In this embodiment, it is set to 0.3, which can smooth the network handover trend and reduce the risk of handover jitter when the reliability of the prediction results is high.

[0044] By quantitatively analyzing the fluctuation amplitude and dispersion of the predicted quality index over continuous time, the system can objectively assess the inherent uncertainty of the neural network prediction results. Based on this assessment, the generated credible quantification factor enables the system to distinguish between high-confidence steady-state predictions and transient fluctuations caused by random disturbances. Accordingly, it performs adaptive weighted smoothing or suppression correction on the original prediction index, transforming the time series statistical characteristics in the prediction model output into actionable decision information. This significantly enhances the reliability of the final decision index in representing the evolution trend of the real network state, effectively suppresses unnecessary switching oscillations caused by instantaneous model misjudgments or sudden channel noise, and improves the robustness and accuracy of the overall switching strategy.

[0045] Specifically, the decision control module includes: An interval determination unit is used to determine whether the cellular network, the wireless local area network, or the dedicated network meets the handover triggering conditions based on the decision index falling within a preset decision interval, and to generate a corresponding candidate network set. The sorting and comparison unit, which is connected to the interval determination unit, is used to sort the corresponding decision indices in the candidate network set from largest to smallest, and mark the first decision index as the primary index and the second decision index as the alternative index. The strategy selection unit, connected to the sorting comparison unit, is used to combine the current service type and the difference between the primary index and the alternative index to generate a single-network switching strategy, a multi-network concurrent strategy, or a pre-switching strategy to obtain the network switching strategy.

[0046] Specifically, the strategy selection unit includes: A single-network handover generation subunit is used to generate the single-network handover strategy when the primary selection index is higher than a preset first handover threshold and the difference between the primary selection index and the alternative index is greater than a preset difference threshold. A multi-network concurrent generation subunit is used to generate the multi-network concurrent strategy when both the primary index and the alternative index are higher than a preset second concurrency threshold, and the difference is less than or equal to the preset difference threshold. The pre-switching generation subunit is used to generate the pre-switching strategy when the primary selection index is higher than the preset third pre-switching threshold and the latency sensitivity level of the current service meets the preset pre-switching conditions.

[0047] Specifically, the strategy selection unit further includes: The indicator extraction subunit is used to extract the maximum allowable latency of a service from the service data packets at the terminal side. The level calculation subunit, which is connected to the index extraction subunit, is used to calculate the corresponding latency sensitivity level based on the maximum allowable latency and according to the preset level classification rules.

[0048] In this embodiment, the preset level classification rules include classifying services with a maximum allowed latency lower than a preset first latency threshold as high latency sensitive level, classifying services with a maximum allowed latency between the preset first latency threshold and a preset second latency threshold as medium latency sensitive level, and classifying services with a maximum allowed latency higher than a preset second latency threshold as low latency sensitive level.

[0049] In this embodiment, the interval determination unit presets four consecutive decision index intervals: high-quality interval, good interval, medium interval, and poor interval. The corresponding interval boundary thresholds are set according to long-term network quality statistics and service experience requirements. For example, the high-quality interval is [0.7, 1.0], the good interval is [0.4, 0.7], the medium interval is [0.2, 0.4], and the poor interval is [0, 0.2]. The interval determination unit compares the decision indices of cellular networks, wireless LANs, and dedicated networks with these preset intervals. Networks whose decision indices fall into the high-quality or good intervals are determined to meet the handover triggering conditions and added to the candidate network set. If the decision index of a network falls into the poor interval, it is excluded from the candidate set to avoid handover to a network with poor quality.

[0050] In this embodiment, after obtaining the candidate network set, the ranking comparison unit sorts the decision indices of all networks in the set in descending order. After sorting, the index ranked first is marked as the primary index, and its corresponding network is marked as the primary network; the index ranked second is marked as the candidate index, and its corresponding network is marked as the candidate network. If there is only one network in the candidate network set, the candidate index is recorded as 0, and the candidate network is recorded as invalid.

[0051] In this embodiment, the strategy selection unit generates a final strategy based on the ranking results, index differences, and service types. Specifically, the single-network switching generation subunit determines whether the following conditions are met: the primary selection index is higher than a preset first switching threshold (e.g., 0.6), and the difference between the primary and alternative selection indices is greater than a preset difference threshold (e.g., 0.15). If both conditions are met, a single-network switching strategy to switch to the primary network is generated. The multi-network concurrent generation subunit determines whether another set of conditions is met: both the primary and alternative selection indices are higher than a preset second concurrency threshold (e.g., 0.5), and the difference between them is less than or equal to the aforementioned preset difference threshold (e.g., 0.15). If these conditions are met, a multi-network concurrent strategy connecting both the primary and alternative networks is generated, and service traffic is distributed across the two links using a link aggregation protocol such as MPTCP.

[0052] In this embodiment, the pre-switching generation subunit is specifically designed for high-latency-sensitive services such as real-time voice. Its judgment logic is as follows: when the primary selection index is higher than a preset third pre-switching threshold (e.g., 0.55, slightly lower than the single-network handover threshold to enable early action), and the current service is determined by the level calculation subunit to be of a high-latency-sensitive level, a pre-switching strategy is generated. This strategy instructs the firmware execution module to establish a control plane connection with the primary network or complete link-layer pre-association operations in advance, while the user data stream continues to be transmitted on the original network. Once the decision index of the current serving network falls out of the predefined good range, the system immediately performs a seamless handover of the data stream.

[0053] In this embodiment, the indicator extraction subunit extracts or infers the maximum allowed end-to-end latency of a service from data packets flowing through the terminal using deep packet inspection technology or parsing service signaling such as SIP protocol session descriptions. For example, VoIP services typically have a latency of 150 milliseconds. The level calculation subunit calculates according to preset level classification rules, where the preset first latency threshold (e.g., 50 milliseconds) and preset second latency threshold (e.g., 200 milliseconds) are based on the latency tolerance of typical interactive services and backend services. Accordingly, services with a maximum allowed latency of less than 50 milliseconds, such as cloud gaming or remote control, are classified as high latency sensitive; services with latency between 50 and 200 milliseconds, such as video conferencing or ordinary voice calls, are classified as medium latency sensitive; and services with latency greater than 200 milliseconds, such as file downloads or software updates, are classified as low latency sensitive. The result of this service latency sensitivity level determination is directly used to trigger the pre-handover strategy and can dynamically fine-tune other types of handover thresholds, for example, using a more aggressive pre-handover threshold for high latency sensitive services.

[0054] The sources or design basis of the various thresholds involved in this embodiment can be summarized as follows: The decision index intervals, such as the boundary thresholds for excellent, good, medium, and poor, are determined based on statistical analysis of historical network quality data and mapping the correspondence between different score intervals and actual service experiences, such as smooth, acceptable, and laggy. The trigger thresholds for various handover strategies, such as the first, second, and third handover thresholds and the difference threshold, are determined based on the aforementioned decision intervals, combined with the handover objectives of each strategy, such as stability, speed, or concurrency, through joint optimization via system simulation and real-world scenario testing. The thresholds for classifying service latency sensitivity levels, such as the first and second latency thresholds, are directly referenced or adapted to relevant communication standards and typical services, such as 3GPP or ITU-T recommendations and end-to-end latency requirements defined by mainstream applications. All thresholds will ultimately be fine-tuned and solidified in a real network environment through a closed-loop feedback mechanism and A / B testing, thus forming a self-consistent and adaptable threshold system.

[0055] By constructing a dynamically adjustable hierarchical decision-making framework, the system collaboratively matches and quantifies the modified decision index reflecting the overall network status with the latency tolerance determined by the inherent attributes of service data packets and the target differences of different handover actions (such as single-network handover, concurrent, or pre-handover). This framework initially screens the network quality status through preset decision intervals, then evaluates the relative advantages between networks based on index ranking and difference relationships, and finally precisely triggers the most suitable handover strategy by combining the transmission characteristics of the service itself. This enables a tight coupling between the quantitative assessment of network status, the objective constraints of service requirements, and the execution objectives of handover actions, thereby significantly improving the targeting, accuracy, and reliability of network handover strategies.

[0056] Specifically, the firmware execution module includes: The policy parsing unit is used to parse the network handover policy to determine the target network, handover type, and handover execution priority; The instruction generation unit is used to generate a sequence of hardware control instructions that matches the network communication protocol stack of the target network based on the target network, the switching type and the priority, so as to obtain the protocol stack control instructions. The driver execution unit is used to call the dedicated driver interface of the hardware acceleration unit to execute the protocol stack control instructions, so as to control the corresponding radio frequency front-end and baseband processing unit on the terminal side to complete the switching operation to the target network.

[0057] Specifically, the instruction generation unit includes: The protocol matching subunit is used to determine the corresponding network communication protocol stack type according to the network type of the target network, and select a control instruction template that matches the network communication protocol stack type from a preset protocol instruction template set; The parameter filling subunit is used to fill the configurable parameters in the control instruction template based on the priority of the switching execution and the current network operating status of the terminal, so as to generate specific protocol stack control instructions. The timing arrangement subunit is used to sort and combine the protocol stack control instructions according to the dependencies between them, so as to form the protocol stack control instructions that can be executed sequentially.

[0058] In this embodiment, the policy parsing unit is responsible for parsing the received network handover policy. Specifically, it extracts the target network identifier, handover type, and handover execution priority set by the decision control module from the policy. The target network identifier, such as the Cell Global Identifier (CGI) for cellular networks, the Base Station Service Set Identifier (BSSID) for wireless LANs, or the Logical Network Number (LN) for private networks, is used to uniquely identify the network object to be handed over or prepared for handover on the terminal side. The handover type explicitly indicates whether the operation is a hard handover, soft handover, or pre-handover. The execution priority characterizes the urgency of the handover task in system scheduling; higher-priority handover tasks will be allocated tighter time constraints and more system resources.

[0059] Based on the output of the policy parsing unit, the instruction generation unit generates protocol stack control instructions that can directly drive the hardware. Its internal protocol matching subunit first selects the corresponding basic instruction template from a pre-set set of protocol instruction templates according to the type of the target network. This template set encompasses standard protocol control procedures for various standards such as cellular networks, wireless LANs, and private networks. For example, the template for Wi-Fi handover includes a framework of standard steps such as RF tuning, link layer association authentication, and network layer address configuration.

[0060] Subsequently, the parameter filling subunit fills in specific parameters into the selected instruction template based on the specific target network identifier, handover priority, and the terminal's real-time network operating status. Network operating status information includes currently active network connections, the operating frequency of the RF front-end, and the computational load of the baseband processing unit. For example, when the handover priority is marked as high, the system sets the timeout parameter for the associated request to a shorter 100 milliseconds instead of the default 300 milliseconds, thereby speeding up the handover response.

[0061] Next, the timing orchestration subunit sorts and integrates the filled instructions according to the logical dependencies between the underlying instructions, forming a sequence of instructions that can be executed sequentially. These dependencies are inherent protocol specifications; for example, it must be ensured that instructions configuring the RF front-end operating channel are executed before instructions sending link layer connection requests.

[0062] Finally, the driver execution unit calls the driver interface of the terminal-side hardware acceleration unit, such as the NPU or a dedicated communication processor, to issue and execute the orchestrated instruction sequence in stages. Especially when executing the pre-handover strategy, this unit first executes the portion of the instruction sequence used for network preparation, such as completing link-layer authentication of the target network, but does not activate its data forwarding path. Once the final handover trigger condition is met, the driver execution unit quickly executes the remaining path handover instructions, thereby completing the data flow migration in a very short time and achieving fast and stable network handover.

[0063] By transforming the network handover strategy generated by upper-layer intelligent decision-making into a sequence of hardware control instructions that strictly adheres to the target network communication protocol stack specifications and the timing dependencies of underlying hardware operations, a precise and efficient mapping from decision logic to physical layer operations is achieved. The system dynamically adjusts the execution time limits and resource scheduling in the instruction sequence based on the priority parameters in the strategy, and adapts specific radio frequency and baseband configuration parameters based on the real-time network status of the terminal. This ensures that the upper-layer strategy intent, the inherent constraints of the communication protocol, and the physical characteristics of the hardware units work in synergy. This guarantees that, for different modes such as pre-handover and hard handover, RF front-end reconfiguration, link reconstruction, and data path switching can be completed with optimal timing while meeting the protocol stack state machine transition requirements. Ultimately, this achieves millisecond-level low-latency and stable execution of the entire intelligent handover system's decision-making closed loop.

[0064] Please see Figure 4 As shown, this is a flowchart of the intelligent handover method for triple-play convergence based on a lightweight neural network in this embodiment. Furthermore, this embodiment also provides an intelligent handover method for triple-play convergence based on a lightweight neural network, including: Real-time collection of network quality parameters for cellular networks, wireless LANs, and dedicated networks at the terminal side; Based on the network quality parameters, a feature input containing historical time series is constructed and input into a preset lightweight neural network model to predict the network quality of the cellular network, the wireless local area network and the dedicated network in the future time window, and output the corresponding predicted quality index. Based on the change characteristics of the prediction quality index over several consecutive prediction periods, a reliable quantification factor is generated to reflect the prediction stability. Based on the change trend determination result of the reliable quantification factor within the future time window, the prediction quality index is jointly corrected or suppressed to obtain a decision index for switching determination. Based on the numerical range distribution and relative ranking relationship of the decision index, and combined with the current service, a network switching strategy is generated. Based on the network switching strategy, corresponding protocol stack control instructions are generated and executed by the hardware acceleration unit to complete the switching between the cellular network, the wireless local area network, and the private network.

[0065] By transforming multidimensional heterogeneous network parameters, which reflect the physical and transport layer states of wireless channels and are collected in real time at the edge, into structured features that characterize their historical temporal evolution, and using a lightweight neural network model to mine the inherent coupling and transmission relationships of these parameters in the continuous time dimension, the system predicts their future state. The system then performs quantitative evaluation and adaptive correction based on statistical stability on the prediction results. Finally, based on the corrected decision index and the inherent attributes of the service, it generates and executes handover instructions that strictly match the underlying communication protocol stack and hardware operations. This achieves an end-to-end closed loop from multi-source network data perception, temporal correlation feature extraction, uncertainty quantification processing, to precise strategy generation and execution, significantly improving the timeliness, accuracy, and service adaptability of network handover in complex mobile scenarios.

[0066] The network quality parameter collection and processing involved in the embodiments of the present invention are all completed locally on the user terminal. All data does not need to be uploaded to the cloud, which fully protects user privacy and data security and does not contain any data processing rules that violate laws, social ethics or harm public interests.

[0067] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A three-network converged intelligent handover system based on a lightweight neural network, characterized in that, include: The data acquisition module is used to collect network quality parameters of cellular networks, wireless LANs and private networks in real time on the terminal side. The prediction module, which is a hardware acceleration unit deployed on the terminal side, is used to construct a feature input of historical time series based on the network quality parameters, and input it into a preset lightweight neural network model to predict the network quality of the cellular network, the wireless local area network and the dedicated network in the future time window, and output the corresponding predicted quality index. The credibility correction module is used to generate a credibility quantification factor to reflect the prediction stability based on the change characteristics of the prediction quality index over several consecutive prediction periods, and to perform joint correction or suppression correction on the prediction quality index based on the change trend determination result of the credibility quantification factor within the future time window, so as to obtain a decision index for switching determination. The decision control module is used to generate network switching strategies based on the numerical range distribution and relative ranking relationship of the decision index, combined with the current service. The firmware execution module is used to generate corresponding protocol stack control instructions based on the network switching strategy, and execute the protocol stack control instructions through the hardware acceleration unit to complete the switching between the cellular network, the wireless local area network and the private network.

2. The intelligent handover system for triple-play convergence based on lightweight neural networks according to claim 1, characterized in that, The data acquisition module includes: The parameter acquisition unit is used to acquire signal strength, round-trip time, available bandwidth and packet loss rate from the driver layer of the cellular network, the wireless local area network and the dedicated network, respectively. The time alignment unit is used to perform timestamp alignment and sampling period unification processing on the signal strength, round-trip delay, available bandwidth and packet loss rate from different network sources to form the network quality parameters.

3. The intelligent handover system for triple-play convergence based on lightweight neural networks according to claim 2, characterized in that, The prediction module includes: The feature construction unit is used to construct the network quality parameters into a feature matrix containing multiple historical time points based on a preset time sliding window; The model calculation unit is used to call the lightweight neural network model, calculate the feature matrix, and output the predicted quality index corresponding to the cellular network, wireless local area network and dedicated network in the future time window.

4. The intelligent handover system for triple-play convergence based on lightweight neural networks according to claim 3, characterized in that, The credibility correction module includes: A fluctuation analysis unit is used to calculate the fluctuation amplitude and dispersion of the prediction quality index within a consecutive preset number of prediction periods; A credibility calculation unit is used to generate the corresponding credibility quantification factor based on the fluctuation amplitude and dispersion. A correction control unit is used to jointly correct the prediction quality index when the confidence quantification factor meets a preset stability condition, or to suppress the prediction quality index when the stability condition is not met, so as to obtain the decision index.

5. The intelligent handover system for triple-play convergence based on lightweight neural networks according to claim 4, characterized in that, The decision control module includes: An interval determination unit is used to determine whether the cellular network, the wireless local area network, or the dedicated network meets the handover triggering conditions based on the decision index falling within a preset decision interval, and to generate a corresponding candidate network set. The sorting and comparison unit is used to sort the corresponding decision indices in the candidate network set from largest to smallest, and mark the first decision index as the primary index and the second decision index as the alternative index. The strategy selection unit is used to combine the current service type and the difference between the primary index and the alternative index to generate a single-network switching strategy, a multi-network concurrent strategy, or a pre-switching strategy to obtain the network switching strategy.

6. The intelligent handover system for triple-play convergence based on lightweight neural networks according to claim 5, characterized in that, The strategy selection unit includes: A single-network handover generation subunit is used to generate the single-network handover strategy when the primary selection index is higher than a preset first handover threshold and the difference between the primary selection index and the alternative index is greater than a preset difference threshold. A multi-network concurrent generation subunit is used to generate the multi-network concurrent strategy when both the primary index and the alternative index are higher than a preset second concurrency threshold, and the difference is less than or equal to the preset difference threshold. The pre-switching generation subunit is used to generate the pre-switching strategy when the primary selection index is higher than the preset third pre-switching threshold and the latency sensitivity level of the current service meets the preset pre-switching conditions.

7. The intelligent handover system for triple-play convergence based on lightweight neural networks according to claim 6, characterized in that, The strategy selection unit further includes: The indicator extraction subunit is used to extract the maximum allowable latency of a service from the service data packets at the terminal side. The level calculation subunit is used to calculate the corresponding latency sensitivity level based on the maximum allowable latency and according to the preset level classification rules.

8. The intelligent handover system for triple-play convergence based on lightweight neural networks according to claim 7, characterized in that, The firmware execution module includes: The policy parsing unit is used to parse the network handover policy to determine the target network, handover type, and handover execution priority; The instruction generation unit is used to generate a sequence of hardware control instructions that matches the network communication protocol stack of the target network based on the target network, the switching type and the priority, so as to obtain the protocol stack control instructions. The driver execution unit is used to call the dedicated driver interface of the hardware acceleration unit to execute the protocol stack control instructions, so as to control the corresponding radio frequency front-end and baseband processing unit on the terminal side to complete the switching operation to the target network.

9. The intelligent handover system for triple-play convergence based on lightweight neural networks according to claim 8, characterized in that, The instruction generation unit includes: The protocol matching subunit is used to determine the corresponding network communication protocol stack type according to the network type of the target network, and select a control instruction template that matches the network communication protocol stack type from a preset protocol instruction template set; The parameter filling subunit is used to fill the configurable parameters in the control instruction template based on the priority of the switching execution and the current network operating status of the terminal, so as to generate specific protocol stack control instructions. The timing arrangement subunit is used to sort and combine the protocol stack control instructions according to the dependencies between them, so as to form the protocol stack control instructions that can be executed sequentially.

10. A method for intelligent handover of three-network convergence based on lightweight neural networks, applied to the intelligent handover system for three-network convergence based on lightweight neural networks as described in any one of claims 1-9, characterized in that, include: Real-time collection of network quality parameters for cellular networks, wireless LANs, and dedicated networks at the terminal side; Based on the network quality parameters, a feature input containing historical time series is constructed and input into a preset lightweight neural network model to predict the network quality of the cellular network, the wireless local area network and the dedicated network in the future time window, and output the corresponding predicted quality index. Based on the change characteristics of the prediction quality index over several consecutive prediction periods, a reliable quantification factor is generated to reflect the prediction stability. Based on the change trend determination result of the reliable quantification factor within the future time window, the prediction quality index is jointly corrected or suppressed to obtain a decision index for switching determination. Based on the numerical range distribution and relative ranking relationship of the decision index, and combined with the current service, a network switching strategy is generated. Based on the network switching strategy, corresponding protocol stack control instructions are generated and executed by the hardware acceleration unit to complete the switching between the cellular network, the wireless local area network, and the private network.

Images