Method and system for optimizing inter-system handover decision in heterogeneous wireless networks

Abstract

The present application relates to the technical field of wireless network communication, in particular to a heterogeneous wireless network cross-system handover decision optimization method and system, the steps comprising: collecting five-layer cross-layer attribute vectors of candidate networks at a preset period to construct a multi-dimensional decision attribute matrix; performing short-term prediction on the signal quality of the current service network through a GM(1,1) grey prediction model to trigger the handover process in advance; performing admission filtering on the candidate networks based on a service QoS demand vector; using a deep Q network to adaptively output the optimal weight vector of each decision attribute; constructing a handover cost penalty vector through a handover fatigue factor and a predicted residence time; calculating a comprehensive score based on the improved TOPSIS method by fusing the weight vector and the penalty vector, and introducing a retention threshold to make a handover decision; and continuously optimizing the weight strategy through reward feedback after handover. The present application can effectively reduce the service interruption probability, improve the handover decision accuracy in multiple scenarios, and reduce the overall handover cost.

Description

Technical Field

[0001] This invention relates to the field of wireless network communication technology, and more specifically to a method and system for optimizing cross-system handover decisions in heterogeneous wireless networks. Background Technology

[0002] With the rapid development of mobile communication technology, multiple wireless access technologies such as 4G LTE, 5G NR, Wi-Fi, and WiMAX are deployed simultaneously, forming a heterogeneous wireless network environment with varying coverage, transmission rates, and service capabilities. In this context, user terminals need to frequently switch between different network standards while on the move to maintain a continuous and stable connection quality. However, different networks differ significantly in signal coverage, resource management, and protocol stack design, making cross-system handover far more complex than intra-standard handover. How to make accurate and efficient handover decisions in a multi-network coexistence environment has become an important research topic in the field of heterogeneous networks.

[0003] Most existing handover decision-making methods are based on setting fixed trigger thresholds using real-time measurements of a single network attribute, which has two major drawbacks: First, relying solely on the measurement value at the current moment for handover judgment makes it impossible to predict short-term changes in signal quality, resulting in delayed handover timing. In scenarios where the signal deteriorates rapidly, the handover process is often triggered only after service interruption has occurred. Second, fixed threshold strategies are extremely sensitive to signal measurement noise, and are prone to causing a "ping-pong effect" when the signal quality fluctuates around the threshold, leading to repeated handovers between adjacent networks by the terminal, resulting in a large amount of unnecessary handover overhead and a decline in user experience.

[0004] Furthermore, in the field of multi-attribute comprehensive handover decision-making, existing solutions typically employ the Analytic Hierarchy Process (AHP) or expert experience to pre-define fixed weights for each attribute, failing to adaptively adjust to dynamic changes in terminal mobility, service type, and network load. In addition, existing methods generally neglect the expected availability duration of candidate networks and the impact of recent handover behavior on the current decision when evaluating candidate networks. This leads to the system frequently selecting edge coverage areas that the terminal is about to traverse as handover targets, or repeatedly handing over to the same network within a short period, resulting in high handover costs and limited actual connection quality improvement. Therefore, there is an urgent need for a heterogeneous network handover decision optimization method that can comprehensively perceive service needs, adaptively adjust decision weights, and consider handover costs. Summary of the Invention

[0005] In view of the above-mentioned shortcomings of the prior art, the present invention provides a method and system for optimizing cross-system handover decisions in heterogeneous wireless networks, which can effectively solve the problems mentioned in the prior art.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] This invention provides a method for optimizing cross-system handover decisions in heterogeneous wireless networks, comprising the following steps:

[0008] S100. Collect the five-layer cross-layer attribute vectors of all detectable candidate wireless networks at a preset period to construct a multi-dimensional decision attribute matrix; the five-layer cross-layer attributes include physical layer attributes, link layer attributes, network layer attributes, service layer attributes and user layer attributes.

[0009] S200. Establish a GM(1,1) grey prediction model for the signal quality time series of the current service network, and output the signal quality prediction sequence within the preset prediction window; when the signal quality prediction value at any time in the prediction sequence falls below the handover trigger threshold, output a trigger signal to start the handover decision process in advance.

[0010] S300. Identify the currently active service type and obtain the corresponding QoS requirement vector. Perform an admission test on the candidate network set based on the QoS requirement vector to generate a subset of candidate networks that meet the QoS constraints.

[0011] S400: Extract the current system state feature vector and input it into the adaptive weight engine based on the deep Q network, and output the adaptive optimal weight vector of each attribute of the current switching decision.

[0012] S500: Calculate the handover fatigue factor and predicted dwell time for each candidate network, and jointly construct a handover cost penalty vector; the handover fatigue factor is an exponential decay function of the number of handovers to the candidate network within a preset time window; the predicted dwell time is estimated based on the terminal's current moving speed and the candidate network's coverage radius.

[0013] S600: Based on the improved TOPSIS method, the decision attribute matrix is ​​weighted using the weight vector of S400, and the switching cost penalty vector of S500 is introduced to correct the comprehensive score. The candidate network with the highest comprehensive score and exceeding the preset maintenance threshold is selected as the optimal switching target; otherwise, the current network connection is maintained.

[0014] S700: After performing the handover operation, collect the actual QoS indicators after the handover, calculate the reward value and update the experience replay buffer of the deep Q network, and periodically trigger network parameter updates to continuously optimize the weight output strategy.

[0015] Furthermore, in step S100, the physical layer attributes include received signal strength RSS, signal-to-noise ratio SINR, and bit error rate BER; the link layer attributes include packet loss rate and round-trip time RTT; the network layer attributes include available bandwidth and current load rate; the service layer attributes include currently active service type code and QoS requirement vector; and the user layer attributes include terminal moving speed, moving direction angle, number of handovers within a preset historical time window, and usage cost.

[0016] Furthermore, the process of establishing the GM(1,1) grey prediction model in step S200 includes:

[0017] Construct an accumulated generation sequence using the K most recent signal quality samples; estimate the parameters of the grey differential equation using the least squares method; output the future... The signal quality prediction sequence and its confidence interval within a time period; when the confidence level of the prediction result is lower than the preset confidence threshold, the historical data sampling window is expanded and the prediction is remodeled.

[0018] Furthermore, the conditions for the admission test in step S300 are as follows:

[0019] The available bandwidth of the candidate network is not lower than the minimum bandwidth requirement of the service, the round-trip latency is not higher than the maximum latency tolerance of the service, the packet loss rate is not higher than the maximum packet loss rate tolerance of the service, and the cost of use is not higher than the maximum cost constraint in the service QoS requirement vector.

[0020] When the standard admission test results in an empty candidate network subset, it automatically downgrades to a looser filtering condition that only tests bandwidth and latency constraints to regenerate the candidate network subset.

[0021] Furthermore, the reinforcement learning framework of the deep Q-network in step S400 is configured as follows:

[0022] The state space is the current system state feature vector, which includes at least signal quality statistical features, bandwidth distribution features, moving speed, historical handover count, service type, and historical QoS default rate;

[0023] The action space is the weight adjustment amount of each decision attribute. The output layer is normalized by the Softmax activation function to satisfy the constraint that the sum of the weights is 1 and non-negative.

[0024] The reward function is the weighted difference between the QoS satisfaction level and the handover cost after the handover.

[0025] Furthermore, the calculation formula for switching the fatigue factor in step S500 is as follows:

[0026] ;

[0027] in, This represents the number of times the candidate network will be switched within a preset time window. The attenuation coefficient;

[0028] The formula for calculating the predicted stay time is:

[0029] ;

[0030] in, The candidate network coverage radius, The vertical distance from the terminal's current location to the network coverage center is represented by the vertical line connecting the terminal's current location. For the terminal's moving speed;

[0031] When the predicted dwell time is lower than the minimum dwell time threshold, an additional penalty coefficient is applied to the overall score of the candidate network.

[0032] Furthermore, the calculation method for the comprehensive score in step S600 is as follows:

[0033] ;

[0034] in, For comprehensive scoring, and Candidate networks The weighted Euclidean distance to the positive and negative ideal solutions. To switch fatigue factors, To predict the length of stay, This is the minimum dwell time threshold;

[0035] when At that time, the decision is to maintain the current network connection and not perform a handover. The candidate network with the highest overall score. The preset hold threshold is used.

[0036] A cross-system handover decision optimization system for heterogeneous wireless networks includes:

[0037] The multi-layer attribute acquisition module is used to periodically acquire five-layer cross-layer attribute vectors of all detectable candidate wireless networks and construct a multi-dimensional decision attribute matrix; the five-layer cross-layer attributes include physical layer attributes, link layer attributes, network layer attributes, service layer attributes and user layer attributes;

[0038] The signal prediction module is used to establish a GM(1,1) grey prediction model for the signal quality time series of the current service network and output the signal quality prediction sequence within a preset prediction window. When the signal quality prediction value at any time in the prediction sequence falls below the handover trigger threshold, a trigger signal is output to start the handover decision process in advance.

[0039] The service-aware filtering module is used to identify the currently active service type and obtain the corresponding QoS requirement vector, perform admission verification on the candidate network set based on the QoS requirement vector, and generate a subset of candidate networks that meet the QoS constraints.

[0040] The adaptive weight module is used to extract the current system state feature vector and input it into the deep Q-network-based adaptive weight engine, and output the adaptive optimal weight vector of each attribute of the current switching decision.

[0041] The handover cost evaluation module is used to calculate the handover fatigue factor and predicted dwell time for each candidate network, and jointly construct a handover cost penalty vector; the handover fatigue factor is an exponential decay function of the number of handovers to the candidate network within a preset time window; the predicted dwell time is estimated based on the terminal's current moving speed and the candidate network's coverage radius.

[0042] The decision execution module is used to weight the decision attribute matrix using a weight vector based on the improved TOPSIS method, and introduce a switching cost penalty vector to correct the comprehensive score. The candidate network with the highest comprehensive score that exceeds the preset retention threshold is selected as the optimal switching target; otherwise, the current network connection is maintained.

[0043] The feedback module is used to collect the actual QoS indicators after the handover operation, calculate the reward value and update the experience replay buffer of the deep Q network, and periodically trigger network parameter updates to continuously optimize the weight output strategy.

[0044] The present invention also provides a terminal, including a processor and a storage medium;

[0045] The storage medium is used to store instructions;

[0046] The processor is configured to operate according to the instructions to perform the steps of the heterogeneous wireless network cross-system handover decision optimization method according to any one of claims 1-7.

[0047] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the heterogeneous wireless network cross-system handover decision optimization method.

[0048] The technical solution provided by this invention has the following advantages compared with the known prior art:

[0049] This invention establishes a GM(1,1) grey prediction model based on the signal quality time series of the current serving network. Within the prediction window, it anticipates short-term trends in signal quality changes, enabling the handover decision-making process to be triggered before the signal quality actually falls below the handover threshold. Compared to passive triggering mechanisms based on real-time measurements, this invention advances the handover preparation time by several sampling periods, reserving ample time for candidate network evaluation and handover execution, thereby effectively reducing the probability of service interruption and QoS default rate caused by handover delays.

[0050] This invention introduces a Deep Q Network (DQN) as an adaptive weight engine. It takes a system state vector containing multi-dimensional information such as signal quality statistical characteristics, terminal mobility speed, service type, and historical QoS default rate as input, and outputs the optimal weight vector of each decision attribute online. It continuously updates network parameters by the actual QoS indicators after switching, and can autonomously adjust with the dynamic changes of network environment and user behavior. It can still maintain high decision accuracy in complex scenarios such as high-speed movement and changing service types.

[0051] This invention constructs a handover cost penalty vector by combining two quantitative indicators, handover fatigue factor and predicted dwell time, and integrates it into the improved TOPSIS comprehensive score. This significantly reduces the number of invalid handovers while ensuring the necessity of handover, thereby reducing the user experience loss caused by signaling overhead and handover latency. Attached Figure Description

[0052] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0053] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0054] Figure 2 This is a schematic diagram of the system structure of the present invention. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0056] The present invention will be further described below with reference to embodiments.

[0057] Example:

[0058] Reference Figure 1 The method flow shown and Figure 2 The system architecture shown below will be used to describe in detail the heterogeneous wireless network cross-system handover decision optimization method of the present invention with specific embodiments. This embodiment takes a terminal in a heterogeneous wireless network environment where LTE / NR and Wi-Fi / WiMAX coexist as an example. The terminal simultaneously holds multiple wireless interfaces and can concurrently detect the status information of all accessible networks in the vicinity.

[0059] S100, five-layer cross-layer attribute collection and multi-dimensional decision attribute matrix construction.

[0060] Specifically, the terminal scans all detectable candidate wireless networks through each wireless interface at a preset acquisition period (e.g., 500ms), acquiring five-layer cross-layer attribute vectors to construct a multi-dimensional decision attribute matrix. , where row index i corresponds to the i-th candidate network and column index j corresponds to the j-th decision attribute.

[0061] The specific attributes of the five layers are as follows: Physical layer attributes include Received Signal Strength (RSS), Signal-to-Noise Ratio (SINR), and Bit Error Rate (BER); Link layer attributes include Packet Loss Rate and Round-Trip Time (RTT); Network layer attributes include Available Bandwidth and Current Load Rate; Service layer attributes include Current Active Service Type Codes (such as video streaming, VoIP, file transfer, etc.) and QoS Demand Vectors; User layer attributes include Terminal Mobility Speed, Mobility Direction Angle, Number of Handovers within a Preset Historical Time Window, and Usage Cost of Each Candidate Network.

[0062] The introduction of the above five attributes enables the decision matrix to comprehensively characterize the candidate network's overall performance across multiple dimensions, including wireless transmission quality, network resource status, service adaptability, and user behavior preferences, providing a complete data foundation for subsequent multi-attribute decision-making. Each attribute is normalized after collection. Benefit-type attributes are normalized forward, with higher values ​​being better, such as bandwidth and signal-to-noise ratio; cost-type attributes are normalized backward, with lower values ​​being better, such as latency and packet loss rate. All attributes are mapped to the [0,1] interval to eliminate dimensional differences.

[0063] S200, an early switching triggering mechanism based on GM(1,1) gray prediction.

[0064] Specifically, by making short-term predictions about the signal quality of the current service network, the timing of handover can be determined in advance, thereby avoiding service interruptions or QoS degradation caused by waiting for the signal to actually deteriorate before triggering the handover.

[0065] The original time series is constructed by collecting the signal quality sample values ​​(such as RSS or SINR) of the current service network at the most recent K consecutive time moments. Perform an accumulation generation operation (1-AGO) on it to obtain the generated sequence. ,in (i ranges from 1 to k). The purpose of a single accumulation is to reduce the random fluctuations in the original sequence, making it exhibit an approximately exponential growth pattern, thereby satisfying the modeling assumption of GM(1,1).

[0066] Based on the generated sequence, the following grey differential equation is established:

[0067] ;

[0068] in, The value of the sequence is generated by accumulating it once; The development coefficient reflects the trend of signal quality changes. Indicates attenuation. (This represents enhancement); b is called the gray action, corresponding to the endogenous driving term of signal quality. A sequence is generated by constructing neighbor means. The grey differential equation is transformed into a system of linear equations, and then the parameter estimates are obtained using the least squares method. .

[0069] After obtaining the parameters, solving the differential equation yields the predicted signal quality over the future Tpre time interval. Then, an IAGO (Incremental Incremental Goal) operation is performed to reconcile the predicted sequence. ⁾ Restored to signal quality prediction sequence It also outputs the confidence interval of the prediction result. If the confidence level is lower than the preset confidence threshold (e.g., 90%), the historical sampling window is automatically expanded (e.g., K is increased from 10 to 20) and the model is remodeled to improve the prediction accuracy.

[0070] When the predicted signal quality value at any time in the prediction sequence falls below the pre-set handover trigger threshold (e.g., RSS below -85dBm), a trigger signal is immediately output and the handover decision process is initiated ahead of schedule. Compared to the traditional passive triggering mechanism based on real-time measurements, gray prediction triggering can advance the handover decision timing by several sampling periods, providing ample time for subsequent candidate network evaluation and handover execution, effectively reducing handover latency and the probability of service interruption.

[0071] S300, Service-Aware QoS Admission Filtering.

[0072] Specifically, the terminal identifies the currently active service type (such as real-time video, VoIP, FTP download, etc.) and extracts the corresponding QoS requirement vector from the locally pre-configured service QoS requirement mapping table. This vector contains four indicators: minimum bandwidth requirement, maximum latency tolerance, maximum packet loss rate tolerance, and maximum cost constraint.

[0073] Based on the above QoS requirement vector, an admission test is performed on all candidate networks: the candidate network must simultaneously meet the following constraints: available bandwidth is not lower than the minimum bandwidth requirement of the service, round-trip latency is not higher than the maximum latency tolerance of the service, packet loss rate is not higher than the maximum packet loss rate tolerance of the service, and usage cost is not higher than the maximum cost. Only when all four constraints are met can the candidate network subset be included in the subsequent evaluation.

[0074] Considering that standard admission testing in extreme scenarios may result in an empty subset of candidate networks (i.e., no network can simultaneously satisfy all constraints), this invention designs an automatic degradation mechanism: when the standard filtering result is empty, it automatically degrades to a more lenient filtering condition that only tests the two most critical constraints of bandwidth and latency, and regenerates a subset of candidate networks to ensure the robustness and availability of the switching decision process. This step significantly reduces the computational scale of subsequent multi-attribute decisions by filtering out candidate networks that clearly do not meet business requirements.

[0075] S400 is an adaptive weight engine based on deep Q-networks.

[0076] Specifically, traditional TOPSIS methods typically use manually set fixed weights, which cannot adapt to dynamic changes in terminal mobility, service type, and network environment. This invention introduces a Deep Q-Network (DQN) as an adaptive weight engine, enabling the weight vector to autonomously adjust according to the real-time system state, thereby achieving continuous optimization of the handover decision strategy.

[0077] The three key elements of a reinforcement learning framework are defined as follows:

[0078] State Space: The current system state feature vector s contains at least the following components: the signal quality statistics of the current serving network (mean, variance), the bandwidth distribution characteristics of candidate networks, the terminal's moving speed, the number of historical handovers within a preset historical time window, the encoding of the currently active service type, and the occurrence rate of QoS default events in recent history (i.e., the historical QoS default rate). The comprehensive introduction of the above multi-dimensional features enables the state representation to simultaneously characterize the current state of network quality, user behavior patterns, and service requirements.

[0079] Action Space: The output of the DQN network is a weight adjustment vector Δw for each decision attribute, where each component corresponds to the weight increment of an attribute. The output layer normalizes the current weight values ​​of all attributes using the Softmax activation function, ensuring that the final weight vector w satisfies the constraint that all components are non-negative and their sum is 1. and This ensures the physical rationality of the weights.

[0080] Reward function: After the switch is completed, the system collects the actual QoS indicators and calculates the reward value. ,in The overall satisfaction score of various QoS indicators after the handover (value range [0,1], the higher the better). The handover cost incurred during this handover (including handover delay, signaling overhead, etc.). and These represent the tradeoff coefficients between QoS satisfaction and handover cost. Positive reward signals guide DQN learning to allocate larger weights to directions with high QoS gains, while penalty terms suppress unnecessary frequent handover behavior.

[0081] The DQN network structure is a fully connected deep neural network. The input layer receives the state vector s, which is then transformed nonlinearly through several hidden layers, and the output layer provides the Q-value estimate Q(s,a;θ) for each action. An empirical replay mechanism is used to store historical transition samples, and the training process is stabilized by periodically copying the parameters of the main network from the target network. The specific implementation follows the standard DQN training process.

[0082] S500, calculation of switching fatigue factor and predicted dwell time.

[0083] By quantifying the cost of switching and constructing a penalty vector, the switching history and movement status are incorporated into the final score, fundamentally curbing the ping-pong switching phenomenon.

[0084] (1) Switching fatigue factors Calculation

[0085] Switching fatigue factors The formula for quantitatively describing the frequency with which a terminal switches to a candidate network in recent history is as follows:

[0086] ;

[0087] in, The number of times the terminal performs a handover to the candidate network within a preset time window (e.g., the past 60 seconds); The attenuation coefficient ( ), control the severity of the penalty for fatigue level on the score, The larger the value, the more significant the effect of switching history on suppressing ratings. When hour, This indicates no history switch and no penalty is imposed; with Increase Exponential decay approaching zero means that the overall score for frequently switched targets will be significantly reduced. By introducing an exponential decay form, this invention achieves flexible suppression of the ping-pong effect without completely prohibiting switching to the network.

[0088] (2) Predicted stay time Calculation

[0089] Predicted stay time Estimate the expected dwell time of the terminal after entering the candidate network coverage area according to its current mobility state, avoiding rapid re-handover caused by switching to an edge coverage area that the terminal is about to traverse. The calculation formula is as follows:

[0090] ;

[0091] Where R is the coverage radius of the candidate network base station or access point; denoted as , where is the vertical distance from the terminal's current location to the line connecting the network coverage center and the terminal; is the lateral component between the terminal's movement direction vector and the coverage center's direction vector; v is the terminal's current movement speed. (Numerator) The chord length represents the distance the terminal travels through the candidate network coverage area at a constant speed in the current direction. Dividing this length by the speed v yields the predicted dwell time.

[0092] when Below the preset minimum stay time threshold If this occurs, it indicates that the terminal will soon leave the coverage area of ​​the candidate network after the handover. The corresponding comprehensive score will be subject to an additional penalty coefficient in subsequent steps to prevent meaningless short-term handovers.

[0093] based on and A switching cost penalty vector is jointly constructed for each candidate network, which serves as the basis for further improving the overall TOPSIS score.

[0094] S600, handover decision and hold decision based on improved TOPSIS.

[0095] Specifically, this step improves upon the classic TOPSIS method by incorporating a switching cost penalty mechanism into the final score, thus forming the innovative comprehensive scoring method of this invention.

[0096] First, the normalized decision attribute matrix constructed by S100 is weighted using the adaptive weight vector w output by S400 to obtain the weighted decision matrix. ,in ( (This is the normalized value of the j-th attribute of the i-th candidate network).

[0097] Secondly, determine the ideal solution. and the negative ideal solution : the positive ideal solution takes the optimal value of each attribute in all candidate networks for each attribute component, and the negative ideal solution takes the worst value for each attribute component. Calculate the weighted Euclidean distance from each candidate network to the positive and negative ideal solutions and :

[0098] ;

[0099] ;

[0100] On this basis, introduce the switching fatigue factor and the predicted residence time to correct the classical TOPSIS score, and define the comprehensive score of the candidate network as:

[0101] ;

[0102] where is the classical TOPSIS proximity, with a value range of [0,1]. The larger the value, the closer the comprehensive attributes of the candidate network are to the positive ideal solution;

[0103] is the switching fatigue factor, with a value range of (0,1], which imposes a penalty on frequent switching directions through exponential decay; is the residence time penalty term: when the predicted residence time Td(Ni) is not lower than the shortest threshold Td,min, this term takes 1 and no penalty is imposed on the score; when Td(Ni) < Td,min, this term is less than 1, and an additional penalty inversely proportional to the residence duration is imposed on the candidate network with a short residence;

[0104] The multiplication of the three factors enables the comprehensive score to take into account the multi-attribute comprehensive performance, historical switching cost, and predicted mobility simultaneously.

[0105] Calculate the comprehensive scores of all candidate networks and then select the candidate network with the highest comprehensive score and compare it with the preset retention threshold : If , then is taken as the optimal switching target and the switching operation is executed; if , then it is determined that the current network is still the best choice, the current network connection is maintained, and no switching is executed. The introduction of the retention threshold effectively avoids ineffective switching when the scores of all candidate networks are not significantly better than the current network.

[0106] S700, feedback after switching and DQN online learning.

[0107] Specifically, after the handover operation is completed, the terminal continuously collects the actual QoS indicators after the handover (including measured bandwidth, latency, packet loss rate, etc.), compares them with the QoS requirement vector before the handover, and calculates the QoS satisfaction score. Considering the cost of this switch... The reward value r is calculated according to the reward function defined in S400, and the transition quadruple consisting of (old state s, action a, reward r, new state s') is stored in the experience playback buffer of DQN.

[0108] The system randomly samples small batches of samples from the experience replay buffer at preset triggering intervals (such as after a certain number of switches or timed triggering), and updates the DQN main network parameters by minimizing the temporal difference (TD) error. Target network parameters Then, it is synchronously copied from the main network at a lower frequency (such as every few steps) to ensure training stability.

[0109] Through the aforementioned online continuous learning mechanism, the DQN weight engine can continuously accumulate experience and autonomously optimize weight output strategies in actual deployment environments, enabling the switching decision method to maintain long-term adaptive capability to changes in the network environment and the evolution of user behavior.

[0110] Reference Figure 2 Corresponding to the above method embodiments, the system provided by the present invention consists of the following functional modules:

[0111] The multi-layer attribute acquisition module is used to periodically collect five-layer cross-layer attribute vectors of all detectable candidate wireless networks and construct a multi-dimensional decision attribute matrix. The five-layer cross-layer attributes include physical layer attributes, link layer attributes, network layer attributes, service layer attributes, and user layer attributes.

[0112] The signal prediction module is used to establish a GM(1,1) grey prediction model for the signal quality time series of the current service network and output the signal quality prediction sequence within the preset prediction window. When the signal quality prediction value at any time in the prediction sequence falls below the handover trigger threshold, a trigger signal is output to start the handover decision process in advance.

[0113] The service-aware filtering module is used to identify the currently active service types and obtain the corresponding QoS requirement vectors. Based on the QoS requirement vectors, it performs admission checks on the candidate network set and generates a subset of candidate networks that meet the QoS constraints.

[0114] The adaptive weight module is used to extract the current system state feature vector and input it into the deep Q-network-based adaptive weight engine, and output the adaptive optimal weight vector of each attribute of the current switching decision.

[0115] The handover cost assessment module is used to calculate the handover fatigue factor and predicted dwell time for each candidate network, and jointly construct the handover cost penalty vector. The handover fatigue factor is an exponential decay function of the number of handovers to the candidate network within a preset time window. The predicted dwell time is estimated based on the terminal's current moving speed and the coverage radius of the candidate network.

[0116] The decision execution module is used to weight the decision attribute matrix using a weight vector based on the improved TOPSIS method, and introduce a switching cost penalty vector to correct the comprehensive score. The candidate network with the highest comprehensive score that exceeds the preset retention threshold is selected as the optimal switching target; otherwise, the current network connection is maintained.

[0117] The feedback module is used to collect the actual QoS indicators after the handover operation, calculate the reward value and update the experience replay buffer of the deep Q network, and periodically trigger network parameter updates to continuously optimize the weight output strategy.

[0118] The present invention also provides a terminal, including a processor and a storage medium. The storage medium is used to store program instructions for implementing each step of the above-described switching decision optimization method. When executing the instructions, the processor completes the functional logic of each step from S100 to S700. The terminal can be a mobile device with a multi-mode wireless interface, such as a smartphone, an in-vehicle terminal, or a laptop computer. The processor can be a general-purpose CPU or an integrated NPU to accelerate DQN inference computation.

[0119] The present invention also provides a computer-readable storage medium having a computer program stored thereon. When the program is executed by a processor, it implements all the steps of the above-described switching decision optimization method. The storage medium may be flash memory, solid-state drive, optical disk, or other non-volatile memory.

[0120] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions will not cause the essence of the corresponding technical solutions to deviate from the protection scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for optimizing cross-system handover decisions in heterogeneous wireless networks, characterized in that, Includes the following steps: S100. Collect the five-layer cross-layer attribute vectors of all detectable candidate wireless networks at a preset period to construct a multi-dimensional decision attribute matrix; the five-layer cross-layer attributes include physical layer attributes, link layer attributes, network layer attributes, service layer attributes and user layer attributes. S200. Establish a GM(1,1) grey prediction model for the signal quality time series of the current service network, and output the signal quality prediction sequence within the preset prediction window; when the signal quality prediction value at any time in the prediction sequence falls below the handover trigger threshold, output a trigger signal to start the handover decision process in advance. S300. Identify the currently active service type and obtain the corresponding QoS requirement vector. Perform an admission test on the candidate network set based on the QoS requirement vector to generate a subset of candidate networks that meet the QoS constraints. S400: Extract the current system state feature vector and input it into the adaptive weight engine based on the deep Q network, and output the adaptive optimal weight vector of each attribute of the current switching decision. S500: Calculate the handover fatigue factor and predicted dwell time for each candidate network, and jointly construct a handover cost penalty vector; the handover fatigue factor is an exponential decay function of the number of handovers to the candidate network within a preset time window; the predicted dwell time is estimated based on the terminal's current moving speed and the candidate network's coverage radius. S600: Based on the improved TOPSIS method, the decision attribute matrix is ​​weighted using the weight vector of S400, and the switching cost penalty vector of S500 is introduced to correct the comprehensive score. The candidate network with the highest comprehensive score and exceeding the preset maintenance threshold is selected as the optimal switching target; otherwise, the current network connection is maintained. S700: After performing the handover operation, collect the actual QoS indicators after the handover, calculate the reward value and update the experience replay buffer of the deep Q network, and periodically trigger network parameter updates to continuously optimize the weight output strategy.

2. The heterogeneous wireless network cross-system handover decision optimization method according to claim 1, characterized in that, In step S100, the physical layer attributes include received signal strength RSS, signal-to-noise ratio SINR, and bit error rate BER. The link layer attributes include packet loss rate and round-trip time (RTT); the network layer attributes include available bandwidth and current load rate; the service layer attributes include currently active service type codes and QoS requirement vectors; and the user layer attributes include terminal movement speed, movement direction angle, number of handovers within a preset historical time window, and usage cost.

3. The heterogeneous wireless network cross-system handover decision optimization method according to claim 1, characterized in that, The process of establishing the GM(1,1) grey prediction model in step S200 includes: Construct an accumulated generation sequence using the K most recent signal quality samples; estimate the parameters of the grey differential equation using the least squares method; output the future... The signal quality prediction sequence and its confidence interval within a time period; when the confidence level of the prediction result is lower than the preset confidence threshold, the historical data sampling window is expanded and the prediction is remodeled.

4. The heterogeneous wireless network cross-system handover decision optimization method according to claim 1, characterized in that, The conditions for the admission test in step S300 are as follows: The available bandwidth of the candidate network is not lower than the minimum bandwidth requirement of the service, the round-trip latency is not higher than the maximum latency tolerance of the service, the packet loss rate is not higher than the maximum packet loss rate tolerance of the service, and the cost of use is not higher than the maximum cost constraint in the service QoS requirement vector. When the standard admission test results in an empty candidate network subset, it automatically downgrades to a looser filtering condition that only tests bandwidth and latency constraints to regenerate the candidate network subset.

5. The heterogeneous wireless network cross-system handover decision optimization method according to claim 1, characterized in that, The reinforcement learning framework configuration of the deep Q-network in step S400 is as follows: The state space is the current system state feature vector, which includes at least signal quality statistical features, bandwidth distribution features, moving speed, historical handover count, service type, and historical QoS default rate; The action space is the weight adjustment amount of each decision attribute. The output layer is normalized by the Softmax activation function to satisfy the constraint that the sum of the weights is 1 and non-negative. The reward function is the weighted difference between the QoS satisfaction level and the handover cost after the handover.

6. The heterogeneous wireless network cross-system handover decision optimization method according to claim 1, characterized in that, The formula for calculating the fatigue factor in step S500 is as follows: ； in, This represents the number of times the candidate network will be switched within a preset time window. The attenuation coefficient; The formula for calculating the predicted stay time is: ； in, The candidate network coverage radius, The vertical distance from the terminal's current location to the network coverage center is represented by the vertical distance. For the terminal's moving speed; When the predicted dwell time is lower than the minimum dwell time threshold, an additional penalty coefficient is applied to the overall score of the candidate network.

7. The method for optimizing cross-system handover decisions in heterogeneous wireless networks according to claim 1, characterized in that, The comprehensive score in step S600 is calculated as follows: ； in, For comprehensive scoring, and Candidate networks The weighted Euclidean distance to the positive and negative ideal solutions. To switch fatigue factors, To predict the length of stay, This is the minimum dwell time threshold; when At that time, the decision is to maintain the current network connection and not perform a handover. The candidate network with the highest overall score. This is a preset hold threshold.

8. A cross-system handover decision optimization system for heterogeneous wireless networks, characterized in that, include: The multi-layer attribute acquisition module is used to periodically acquire five-layer cross-layer attribute vectors of all detectable candidate wireless networks and construct a multi-dimensional decision attribute matrix; the five-layer cross-layer attributes include physical layer attributes, link layer attributes, network layer attributes, service layer attributes and user layer attributes; The signal prediction module is used to establish a GM(1,1) grey prediction model for the signal quality time series of the current service network and output the signal quality prediction sequence within a preset prediction window. When the signal quality prediction value at any time in the prediction sequence falls below the handover trigger threshold, a trigger signal is output to start the handover decision process in advance. The service-aware filtering module is used to identify the currently active service type and obtain the corresponding QoS requirement vector, perform admission verification on the candidate network set based on the QoS requirement vector, and generate a subset of candidate networks that meet the QoS constraints. The adaptive weight module is used to extract the current system state feature vector and input it into the deep Q-network-based adaptive weight engine, and output the adaptive optimal weight vector of each attribute of the current switching decision. The handover cost evaluation module is used to calculate the handover fatigue factor and predicted dwell time for each candidate network, and jointly construct a handover cost penalty vector; the handover fatigue factor is an exponential decay function of the number of handovers to the candidate network within a preset time window; the predicted dwell time is estimated based on the terminal's current moving speed and the candidate network's coverage radius. The decision execution module is used to weight the decision attribute matrix using a weight vector based on the improved TOPSIS method, and introduce a switching cost penalty vector to correct the comprehensive score. The candidate network with the highest comprehensive score that exceeds the preset retention threshold is selected as the optimal switching target; otherwise, the current network connection is maintained. The feedback module is used to collect the actual QoS indicators after the handover operation, calculate the reward value and update the experience replay buffer of the deep Q network, and periodically trigger network parameter updates to continuously optimize the weight output strategy.

9. A terminal, comprising a processor and a storage medium; characterized in that: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the steps of the heterogeneous wireless network cross-system handover decision optimization method according to any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the cross-system handover decision optimization method for heterogeneous wireless networks as described in any one of claims 1-7.

Images