A network intelligent switching method and system for a dynamic wireless environment, a terminal and a storage medium

Abstract

The application relates to the technical field of network intelligent switching, and discloses a network intelligent switching method and system for a dynamic wireless environment, a terminal and a storage medium, the method comprising the following steps: acquiring the network state of a user, determining a base station, and confirming a switching behavior according to the determined base station; if the switching behavior exists, acquiring uplink local observation and downlink local observation, and inputting the uplink local observation and the downlink local observation into an uplink-downlink decoupling switching decision network; performing local action value calculation, and jointly processing the calculation results; mapping uplink candidate actions and downlink candidate actions to a target interval, fusing the obtained joint action value and safety score results, performing network switching according to a target switching action of a target action evaluation value, and obtaining a network intelligent switching result. The application can simultaneously realize adaptive smoothing of reward feedback and risk perception screening of candidate switching actions, and improves the success rate and reliability of network intelligent switching.

Description

Technical Field

[0001] This invention relates to the field of intelligent network handover technology, and in particular to an intelligent network handover method, system, terminal, and computer-readable storage medium for dynamic wireless environments. Background Technology

[0002] As 5G mobile communication networks evolve towards high density, multi-layered heterogeneity, and asymmetric uplink and downlink services, heterogeneous cellular networks composed of macro base stations, 5G NR (New Radio) micro base stations, and millimeter-wave micro base stations have become an important architecture for improving system capacity and coverage. Due to differences in transmit power, coverage area, bandwidth resources, and propagation characteristics among different types of base stations, the optimal serving base station for the same user in the uplink and downlink directions may not be the same. Traditional uplink-downlink coupled access methods require users to associate with the same base station in both directions, making it difficult to simultaneously achieve downlink coverage stability and uplink path loss advantages. This can easily lead to problems such as limited uplink coverage, uneven load distribution, and degraded link quality at cell edges.

[0003] Currently, existing technologies propose uplink / downlink decoupling access mechanisms and employ methods such as maximum reference signal received power, fixed thresholds, hysteresis parameters, or load-aware rules for handover control. However, in scenarios involving user mobility, channel fading, and dynamically changing interference relationships, uplink and downlink may experience asynchronous handover, unidirectional handover, or inconsistent target base stations. Traditional handover methods struggle to adaptively optimize based on long-term network benefits and cannot simultaneously balance system throughput, user fairness, and handover stability. Existing deep reinforcement learning or multi-agent reinforcement learning methods can learn adaptive handover strategies through environmental interaction feedback, but their training process is susceptible to instantaneous link fluctuations, abnormal handover events, and abrupt changes in handover penalties, leading to strong randomness in reward signals, unstable value estimation, and fluctuations in policy convergence quality. Existing reward smoothing methods such as fixed window averaging, exponential moving average, or manual threshold pruning typically rely on preset parameters and struggle to adaptively extract representative states based on the true distribution of communication indicator samples. When the reward distribution exhibits multi-peak or non-uniform characteristics, they are prone to insufficient or excessive smoothing. Meanwhile, existing reinforcement learning handover methods mostly select target base stations based on action value or simple action masks, lacking fine-grained evaluation of high-risk actions that may cause handover failure, ping-pong handover, link quality degradation, or load deterioration, making it difficult to effectively suppress potential risks before action execution. Therefore, existing technologies still need improvement and development. Summary of the Invention

[0004] The main objective of this invention is to provide a network intelligent handover method, system, terminal, and storage medium for dynamic wireless environments, aiming to solve the problem that existing technologies cannot simultaneously achieve adaptive smoothing of reward feedback and risk perception screening of candidate handover actions, resulting in low success rate and reliability of network handover.

[0005] To achieve the above objectives, the present invention provides a network intelligent handover method for dynamic wireless environments, the network intelligent handover method for dynamic wireless environments comprising the following steps: The user's network status is obtained, the target base station is determined based on the network status, the uplink target base station and the downlink target base station are obtained, and the handover behavior is confirmed based on the uplink target base station and the downlink target base station to obtain the handover confirmation result; If the handover confirmation result indicates that a handover action exists, then the uplink local observation of the uplink target base station and the downlink local observation of the downlink target base station are obtained based on the handover confirmation result, and the uplink local observation and the downlink local observation are input into the trained uplink-downlink decoupled handover decision network. Local action values ​​are calculated for the uplink local observations and the downlink local observations to obtain the target uplink local action value and the target downlink local action value. The target uplink local action value and the target downlink local action value are then jointly processed to obtain the target joint action value. The uplink candidate actions corresponding to the uplink local observations and the downlink candidate actions corresponding to the downlink local observations are mapped to the target interval to obtain a security score result representing the security confidence of the candidate actions. The joint action value and the security score result are fused to obtain the target action evaluation value. The network switching is performed according to the target switching action of the target action evaluation value to obtain the network intelligent switching result.

[0006] Optionally, the network intelligent handover method for dynamic wireless environments, wherein obtaining the user's network status, determining the target base station based on the network status to obtain the uplink target base station and the downlink target base station, and confirming the handover behavior based on the uplink target base station and the downlink target base station to obtain the handover confirmation result, specifically includes: The network status of the user's location is obtained, and uplink and downlink base stations are associated with the network status based on the base station agent of the area to obtain the base station association result. Based on the base station association result, the base station is determined to obtain the uplink target base station and the downlink target base station. The number of handovers of the uplink target base station and the downlink target base station is statistically processed to obtain handover statistics results. Based on the handover statistics results, the handover behavior of the uplink target base station and the downlink target base station is confirmed to obtain handover confirmation results.

[0007] Optionally, in the network intelligent handover method for dynamic wireless environments, the training process of the uplink / downlink decoupled handover decision network specifically includes: Obtain a network training set and construct an uplink / downlink decoupling decision training network. The network training set includes multiple sets of historical uplink local observations of historical uplink base stations and historical downlink local observations of historical downlink base stations. The uplink / downlink decoupling decision training network includes an uplink / downlink decoupling association network and a spatiotemporal attention residual action filtering network. A set of historical uplink local observations and historical downlink local observations from the network training set are input into the uplink-downlink decoupling decision training network. The uplink-downlink decoupling correlation network encodes the historical uplink local observations and historical downlink local observations respectively to obtain the uplink temporal hidden state and the downlink temporal hidden state. Action values ​​are calculated for the uplink and downlink hidden timing states to obtain uplink local action values ​​and downlink local action values. The uplink local action values ​​and downlink local action values ​​are then combined to obtain a combined action value. Set the original reward, and smooth the original reward according to the subtractive clustering reward smoothing mechanism to obtain the smoothed reward; The spatiotemporal attention residual action filtering network is used to construct features for the historical uplink candidate actions corresponding to the historical uplink local observations and the historical downlink candidate actions corresponding to the historical downlink local observations, so as to obtain uplink risk assessment features and downlink risk assessment features. The uplink risk assessment features and downlink risk assessment features are then mapped to the target interval to obtain uplink security scores and downlink security scores that characterize the security confidence of candidate actions. The combined action value, the uplink security score, and the downlink security score are fused together to obtain a handover action evaluation value. A network intelligent handover action is obtained based on the handover action evaluation value. The network intelligent handover action is executed to obtain an action execution result. The uplink and downlink association of the user is updated based on the execution result. The parameters of the uplink-downlink decoupling decision training network are updated according to the smoothing reward, and the next set of historical uplink local observations and historical downlink local observations are input into the uplink-downlink decoupling decision training network until the training status of the uplink-downlink decoupling decision training network meets the preset conditions, thus obtaining the trained uplink-downlink decoupling decision network.

[0008] Optionally, the network intelligent handover method for dynamic wireless environments, wherein setting the original reward and smoothing the original reward according to a subtractive clustering reward smoothing mechanism to obtain a smoothed reward, specifically includes: Set the initial reward and obtain key communication indicators, and construct a communication status sample based on the key communication indicators; Obtain a sample set, calculate the density potential value of each sample point in the sample set to obtain the calculation result, extract the maximum density potential value in the calculation result, and reduce the samples in the sample set according to the target sample corresponding to the maximum density potential value to obtain a cluster center set. The distance between the communication state sample and each cluster center in the cluster center set is calculated to obtain the distance calculation result. The membership weight is calculated based on the distance calculation result, and the original reward is smoothed based on the membership weight to obtain the smoothed reward.

[0009] Optionally, in the network intelligent handover method for dynamic wireless environments, the setting of the initial reward specifically includes: ; The calculation of the density potential value for each sample point in the sample set is specifically as follows: ; The step of reducing the number of samples in the sample set based on the target sample corresponding to the maximum density potential value specifically involves: ; in, For the original reward, The normalized system transmission rate, The normalized average link quality metric. As a fairness metric used to evaluate load balancing among base stations, This represents the normalized number of handover failures. This represents the normalized number of ping-pong switches. , , , and All are reward weighting coefficients. For the sample density potential, For the first One sample, For the first One sample, To influence the radius of clustering, For the first The residual density potential of each sample after density reduction. The density potential value corresponding to the cluster center. As cluster center, The radius is the density reduction radius.

[0010] Optionally, the network intelligent handover method for dynamic wireless environments, wherein mapping the uplink risk assessment features and the downlink risk assessment features to a target interval to obtain an uplink security score and a downlink security score characterizing the security confidence of candidate actions, further includes: The historical uplink candidate actions and the historical downlink candidate actions are executed to obtain action execution results, and risk labels are constructed based on the execution results; The training loss of the spatiotemporal attention residual action screening network is calculated based on the risk label to obtain the loss value, and the spatiotemporal attention residual action screening network is optimized based on the loss value.

[0011] Optionally, the network intelligent handover method for dynamic wireless environments, wherein updating the parameters of the uplink / downlink decoupling decision training network based on the smoothing reward specifically includes: Obtain the target batch of samples, and calculate the temporal difference target of the target batch of samples based on the smoothing reward; The network training loss is calculated based on the temporal difference objective to obtain the target loss value, and the parameters of the uplink / downlink decoupling decision training network are updated based on the target loss value.

[0012] Optionally, the network intelligent handover method for dynamic wireless environments, wherein the network intelligent handover system for dynamic wireless environments includes: The target base station confirmation module is used to obtain the user's network status, determine the target base station based on the network status, obtain the uplink target base station and the downlink target base station, and confirm the handover behavior based on the uplink target base station and the downlink target base station to obtain the handover confirmation result. The local observation acquisition module is used to acquire uplink local observations of the uplink target base station and downlink local observations of the downlink target base station based on the handover confirmation result if the handover confirmation result indicates that a handover action exists, and input the uplink local observations and the downlink local observations into the trained uplink-downlink decoupled handover decision network. The action value calculation module is used to calculate the local action value of the uplink local observation and the downlink local observation to obtain the target uplink local action value and the target downlink local action value, and to perform joint processing on the target uplink local action value and the target downlink local action value to obtain the target joint action value; The network intelligent handover module is used to map the uplink candidate actions corresponding to the uplink local observation and the downlink candidate actions corresponding to the downlink local observation to the target interval, obtain a security score result representing the security confidence of the candidate actions, fuse the joint action value and the security score result to obtain the target action evaluation value, and perform network handover according to the target handover action of the target action evaluation value to obtain the network intelligent handover result.

[0013] Furthermore, to achieve the above objectives, the present invention also provides a terminal, wherein the terminal includes: a memory, a processor, and a network intelligent handover program for dynamic wireless environments stored in the memory and executable on the processor, wherein when the network intelligent handover program for dynamic wireless environments is executed by the processor, it implements the steps of the network intelligent handover method for dynamic wireless environments as described above.

[0014] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a network intelligent handover program for dynamic wireless environments, and when the network intelligent handover program for dynamic wireless environments is executed by a processor, it implements the steps of the network intelligent handover method for dynamic wireless environments as described above.

[0015] In this invention, the user's network status is acquired, and target base stations are determined based on the network status to obtain uplink and downlink target base stations. Handover behavior confirmation is performed based on the uplink and downlink target base stations to obtain a handover confirmation result. If the handover confirmation result indicates that handover behavior exists, uplink local observations of the uplink target base station and downlink local observations of the downlink target base station are acquired based on the handover confirmation result, and these observations are input into a trained uplink-downlink decoupled handover decision network. Local action values ​​are calculated on the uplink and downlink local observations to obtain target uplink and downlink local action values. These target uplink and downlink local action values ​​are then jointly processed to obtain a target joint action value. Uplink candidate actions corresponding to the uplink local observations and downlink candidate actions corresponding to the downlink local observations are mapped to target intervals to obtain a security score representing the security confidence of the candidate actions. The joint action value and the security score are then fused to obtain a target action evaluation value. Network handover is performed based on the target handover action of the target action evaluation value to obtain a network intelligent handover result. This invention can simultaneously achieve adaptive smoothing of reward feedback and risk perception screening of candidate switching actions, thereby improving the success rate and reliability of intelligent network switching. Attached Figure Description

[0016] Figure 1 This is a flowchart of a preferred embodiment of the network intelligent handover method for dynamic wireless environments of the present invention; Figure 2 This is a schematic diagram of the overall architecture of the network intelligent handover method for dynamic wireless environments according to the present invention; Figure 3 This is a schematic diagram comparing the training convergence curves before and after introducing subtractive clustering reward smoothing in a preferred embodiment of the present invention. Figure 4 This is a schematic diagram comparing the reward volatility before and after introducing subtractive clustering reward smoothing in a preferred embodiment of the present invention; Figure 5 This is a schematic diagram showing the comparison results of different algorithms on switching stability indicators in a preferred embodiment of the present invention; Figure 6 This is a schematic diagram showing the comparison results of different algorithms in terms of comprehensive performance indicators in a preferred embodiment of the present invention; Figure 7 This is a structural diagram of a preferred embodiment of the network intelligent handover system for dynamic wireless environments of the present invention; Figure 8 This is a structural diagram of a preferred embodiment of the terminal of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer and more explicit, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0018] It should be noted that if the embodiments of the present invention involve directional indicators (such as up, down, left, right, front, back, etc.), the directional indicators are only used to explain the relative positional relationship and movement of the components in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indicators will also change accordingly.

[0019] Furthermore, if the embodiments of this invention involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. If the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.

[0020] The preferred embodiment of the network intelligent handover method for dynamic wireless environments described in this invention, such as... Figure 1 As shown, the network intelligent handover method for dynamic wireless environments includes the following steps: Step S10: Obtain the user's network status, determine the base station based on the network status to obtain the uplink target base station and the downlink target base station, and confirm the handover behavior based on the uplink target base station and the downlink target base station to obtain the handover confirmation result.

[0021] Specifically, existing network handover methods cannot simultaneously achieve adaptive smoothing of reward feedback and risk-aware screening of candidate handover actions, resulting in low success rates and reliability of network handover. Therefore, this invention proposes a network intelligent handover method for dynamic wireless environments, constructing an uplink-downlink decoupled heterogeneous cellular network composed of multiple types of base stations and multiple mobile users. The base stations include macro base stations, 5G NR (New Radio) micro base stations, and millimeter-wave micro base stations. Users move within the network coverage area and can select serving base stations for both uplink and downlink. The set of base stations is represented as follows: The user set is represented as ,in, The total number of base stations. This represents the total number of users. For any user... In the time slot The uplink and downlink serving base stations are respectively denoted as , Unlike traditional uplink-downlink coupled access, this invention allows users to associate with different serving base stations in the uplink and downlink directions, i.e. To describe the relationship between users and base stations, uplink and downlink binary correlation variables are defined as follows: ; ; in, and Representing users respectively In the time slot Are the uplink and downlink associated with the base station? When the user's uplink serving base station is At that time, ;otherwise, When the user's downlink serving base station is At that time, ;otherwise, To ensure that each user is served by only one base station in each direction, the system needs to satisfy the following constraints: ; ; The system optimization objective is to improve uplink and downlink transmission performance, user fairness, and handover stability by rationally selecting uplink and downlink correlation matrices under conditions of user mobility, link fading, and dynamic changes in base station load. This optimization objective can be expressed as: ; in, The total system transmission rate As a fairness metric used to evaluate load balancing among base stations, For the average reference signal received power, To switch the number of failures, For the number of times the ping-pong ball is switched, , , , and All are weighting coefficients.

[0022] This invention uses base stations as agents in multi-agent reinforcement learning. Each base station selects distributed actions based on its own local observations, and achieves multi-base station collaboration through centralized value function learning during the training phase. The observation information of each agent includes user location, link quality, reference signal received power, base station load, uplink and downlink correlation of the previous time slot, and historical handover status, etc. The observation variables can be increased or decreased according to the actual network deployment.

[0023] The specific implementation process is as follows: Figure 2 As shown, within each time slot, the system, based on the current network state, collaboratively determines the user's uplink and downlink target base stations by the intelligent agents of each base station. Specifically, it acquires the network state of the user's area, performs uplink and downlink base station association based on the network state using the base station intelligent agents of that area, obtains base station association results, and determines the base stations based on the base station association results to obtain the uplink and downlink target base stations. The base station association results include both uplink and downlink association results. For the user... The uplink target base station corresponding to its uplink association result can be represented as The downlink target base station corresponding to its downlink correlation result can be represented as The base station association result is obtained by the intelligent agent through collaborative decision-making based on information such as local observations, link quality, base station load, and historical handover status.

[0024] Subsequently, handover behavior confirmation is performed based on the uplink target base station and the downlink target base station to obtain a handover confirmation result, wherein the number of handovers is... The expression is: ; in, For users In the time slot Has an uplink handover occurred? For users In the time slot Has a downlink handover occurred?

[0025] Step S20: If the handover confirmation result indicates that a handover action exists, then obtain the uplink local observation of the uplink target base station and the downlink local observation of the downlink target base station based on the handover confirmation result, and input the uplink local observation and the downlink local observation into the trained uplink-downlink decoupled handover decision network.

[0026] Specifically, after obtaining the handover confirmation result, if the handover confirmation result indicates that a handover action has occurred, then the uplink local observations of the uplink target base station and the downlink local observations of the downlink target base station are obtained based on the handover confirmation result. Before inputting the uplink local observations and the downlink local observations into the trained uplink-downlink decoupling handover decision network, a trained uplink-downlink decoupling handover decision network needs to be obtained. The specific training process is as follows: obtain the network training set and construct the uplink-downlink decoupling decision training network. The network training set includes multiple sets of historical uplink local observations of historical uplink base stations and historical downlink local observations of historical downlink base stations. The uplink-downlink decoupling decision training network includes an uplink-downlink decoupling association network and a spatiotemporal attention residual action filtering network. Figure 2 As shown, in order to learn the cooperative handover strategy between different base stations, this invention constructs an uplink-downlink decoupling association network based on GRU-QMIX (Gated Recurrent Unit-based QMIX, a multi-agent reinforcement learning method based on gated recurrent unit timing modeling). This network includes an uplink decision branch (i.e., the uplink GRU-QMIX branch) and a downlink decision branch (i.e., the downlink GRU-QMIX branch), which are used to learn the serving base station selection strategy in the uplink and downlink directions, respectively.

[0027] A set of historical uplink local observations and historical downlink local observations from the network training set are input into the uplink-downlink decoupling decision training network for the agent. The corresponding local observations Hidden state from the previous moment Input the GRU (Gated Recurrent Unit) encoder to obtain the current time-series hidden state. Specifically, the historical uplink local observations and the historical downlink local observations are encoded through the uplink-downlink decoupling network to obtain the uplink temporal hidden state and the downlink temporal hidden state, respectively. Then, the temporal hidden state is input into a local Q-network to obtain the action value of the agent under different candidate actions. Specifically, the action value is calculated for the uplink and downlink temporal hidden states to obtain the uplink local action value and the downlink local action value, where the expressions for the uplink local action value and the downlink local action value are as follows: ; ; in, For intelligent agents In the time slot For upward candidate actions The output local action value (i.e., the uplink local action value). For intelligent agents In the time slot Downlink candidate actions The output local action value (i.e., the downlink local action value). For uplink local Q-network mapping function, This is the downlink local Q-network mapping function. For intelligent agents The temporal hidden state in the uplink direction, For intelligent agents Temporal hidden state in the downlink direction.

[0028] Subsequently, the local Q-values ​​of each agent are input into the QMIX hybrid network, and combined with the global state to obtain the joint action value. Specifically, the uplink local action value and the downlink local action value are jointly processed to obtain the joint action value, the corresponding expression of which is: ; in, For the value of joint operations, For hybrid networks, The local action value of the first base station agent. The local action value of the second base station agent. This represents the global state. Through the above structure, the system can learn the collaborative relationships between multiple base stations during the training phase, and during the execution phase, each base station completes the distributed handover decision based on its own local observations.

[0029] Next, the initial reward is set. This invention incorporates system throughput, link quality, load fairness, and handover stability into the reward design. Therefore, the expression for the initial reward is: ; in, For the original reward, The normalized system transmission rate, The normalized average link quality metric. As a fairness metric used to evaluate load balancing among base stations, This represents the normalized number of handover failures. This represents the normalized number of ping-pong switches. , , , and All of these are reward weighting coefficients. This original reward ensures that the agent, during the learning process, not only pursues higher system transmission performance but also considers user fairness and switching stability, avoiding frequent or high-risk switching behaviors in pursuit of short-term speed improvements.

[0030] Due to the randomness of wireless channel fluctuations, user movement, and handover events, directly using the original reward... This could lead to drastic fluctuations in training rewards. Therefore, this invention introduces a subtractive clustering reward smoothing mechanism during the training phase. Subtractive clustering does not require pre-specifying the number of clusters and can automatically extract representative states based on the density distribution of communication indicator samples, making it suitable for dynamic switching training scenarios with multi-peak reward distributions and significant fluctuations. First, key communication indicators are extracted from the environmental interaction process, and communication state samples are constructed. For a training window or a set of samples in the experience replay pool ,in, For the first sample, This is the second sample. For the first For each sample in the sample set, the density potential value of each sample point is calculated, and the expression corresponding to the calculation result is as follows: ; in, For the sample density potential, For the first One sample, For the first One sample, Let the cluster influence radius be denoted by ; the sample with the highest density potential (i.e., the target sample) is selected as the cluster center, and the density of samples surrounding the target sample is reduced. The corresponding expression is: ; in, For the first The residual density potential of each sample after density reduction. The density potential value corresponding to the cluster center. As cluster center, The density reduction radius is used. This reduction process reduces the likelihood that samples from the neighborhood of the current cluster center will be selected as new cluster centers, thus preventing over-concentration of cluster centers. Repeating this process yields the set of cluster centers. .

[0031] The distance between the communication state sample and each cluster center in the cluster center set is calculated to obtain the distance calculation result, and the corresponding expression is: ; in, Indicates sample For the first The degree of membership of each cluster center This is the smoothing coefficient. A smoothing reward is generated based on the representative reward value corresponding to each cluster center: ; in, For time slots The corresponding smooth reward, The number of cluster centers. For the first The representative reward value corresponding to each cluster center can be determined by the original mean reward of the samples in the neighborhood of that cluster center or the reward of the center sample.

[0032] During the training phase, smooth rewards will be applied. Replacement of original reward The data is written into the experience replay pool and used for subsequent temporal differential target updates. This mechanism allows for the adaptive generation of more stable training feedback based on the distribution characteristics of communication indicator samples, reducing the interference of instantaneous channel fluctuations, anomalous samples, and occasional handover events on value estimation. A comparison of the training convergence curves before and after introducing subtractive clustering reward smoothing is shown below. Figure 3 As shown, Figure 3 The light-colored curve represents the original training reward curve without moving average processing, reflecting the original fluctuations during training; the dark-colored curve represents the reward change trend after 30 rounds of moving average processing; the comparison results of reward volatility before and after introducing subtractive clustering reward smoothing are shown below. Figure 4 As shown, this is used to demonstrate the changes in the rolling coefficient of variation under different training methods.

[0033] It's important to note that subtractive clustering doesn't require repeatedly performing global clustering on all historical samples in every time slot. Instead, it can periodically update cluster centers based on a fixed time window, a small batch of samples from the experience replay pool, or samples accumulated over several training rounds. Between two adjacent clustering updates, new samples only need to be mapped to existing cluster centers to receive a smooth reward. Therefore, this mechanism primarily operates during the training phase and does not significantly increase the time-slot-level computational burden during the online execution phase.

[0034] Subsequently, to further reduce high-risk handover behaviors, this invention introduces a spatio-temporal attentional residual action filtering network, namely the Spatio-Temporal Attentive Residual Filter (STAR-Filter), outside the main decision-making network. STAR corresponds to Spatio-Temporal in spatio-temporal feature modeling, Attentive attention mechanism, and Residual residual structure, respectively. Filter indicates that this module is used to screen candidate handover actions for risk. This spatio-temporal attentional residual action filtering network is used to assess the risk of candidate handover actions, determining whether they may lead to handover failure, ping-pong handover, link quality degradation, or target base station load deterioration. Specifically, the STAR-Filter consists of a feature embedding layer, a spatiotemporal attention module, a residual feature extraction module, and a security score output layer. The feature embedding layer maps risk-related information such as current observations, candidate actions, link quality changes, target base station load changes, and historical handover status to a unified feature space. The spatiotemporal attention module extracts short-term temporal change features during user movement and models spatial interference or load correlation between different base stations. The residual feature extraction module enhances the deep feature representation capability and alleviates gradient decay during training. The security score output layer outputs the security score of candidate actions through a sigmoid function.

[0035] For candidate actions Construct risk assessment features The corresponding expression is: ; in, Indicates current observation information. This indicates the change in link quality before and after the handover. This indicates a change in the target base station load. For historical transition, Switching failed. For ping-pong switching information, specifically, the spatiotemporal attention residual action filtering network is used to construct features for historical uplink candidate actions corresponding to historical uplink local observations and historical downlink candidate actions corresponding to historical downlink local observations, obtaining uplink risk assessment features and downlink risk assessment features; subsequently, the STAR-Filter outputs a safety score for the candidate actions based on the above risk features. The corresponding expression is: ; in, For the action filtering network, Its network parameters, .when The larger the value, the lower the risk of the action; when... The smaller the value, the more likely the action is to cause an unstable handover event. Specifically, the uplink risk assessment features and the downlink risk assessment features are mapped to a target interval (the target interval is [0, 1]) to obtain the uplink security score and downlink security score that characterize the security confidence of the candidate action.

[0036] Subsequently, during the action selection phase, the main GRU-QMIX network outputs candidate action values, and the STAR-Filter outputs the security scores of the candidate actions. By fusing these two values, the final action evaluation value (i.e., the handover action evaluation value) is obtained. Specifically, the joint action value, the uplink security score, and the downlink security score are fused to obtain the handover action evaluation value, with the corresponding expression as follows: ; in, The normalized value of the main network action. The fusion coefficient is used to determine the network intelligent handover action based on the handover action evaluation value. The corresponding expression is: In the early stages of training, the risk labels in the STAR-Filter are insufficient, allowing the value of the main network actions to have a relatively large weight. As training progresses and risk samples accumulate, the role of safety scores in action selection is gradually enhanced. Thus, the system can comprehensively balance communication performance gains and handover risks.

[0037] Subsequently, the network intelligent handover action is processed to obtain the action execution result, and the uplink and downlink association of the user is updated based on the execution result. Specifically, the training samples in the experience replay pool can be represented as follows: After sampling a small batch of samples (i.e., the target batch of samples) from the experience replay pool, the temporal difference objective of the target batch of samples is calculated based on the smoothing reward, and the corresponding expression is: ; in, The value of the joint action output by the main GRU-QMIX network. As a discount factor, This is the final switching action. Based on the stated temporal difference objective, the network training loss is calculated to obtain the target loss value, and the corresponding expression is: ; in, For the time-difference objective, For the value of joint operations, The parameters are updated based on the target loss value. The next set of historical uplink and downlink local observations are then input into the uplink-downlink decoupling decision training network until the training status of the network meets preset conditions (e.g., a preset number of training rounds), resulting in a trained uplink-downlink decoupling decision network.

[0038] Furthermore, to train the spatiotemporal attention residual action selection network, risk labels can be constructed based on the actual switching results within a time window after the action execution. Specifically, the historical uplink candidate actions and the historical downlink candidate actions are processed to obtain the action execution results; then, risk labels are constructed based on the execution results, with the corresponding expression being: ; ; in, This is a risk event indicator variable after the candidate action is executed. It is used when no risk events such as handover failure, ping-pong handover, link quality degradation, or load deterioration occur within the subsequent observation window. Corresponding security label When a risk event occurs Corresponding risk label The training loss of the spatiotemporal attention residual action screening network is calculated based on the risk labels to obtain the loss value. The corresponding expression is: ; The spatiotemporal attention residual action filtering network is optimized based on the loss value. Through this mechanism, the system can perform risk perception and judgment on candidate switching actions before the action is executed, thereby suppressing high-risk switching behaviors.

[0039] Furthermore, to verify the handover decision performance of the trained uplink and uplink decoupled handover decision network, the proposed method SGQ-SM (STAR-Filter-enhanced GRU-QMIX with Smoothing Mechanism) was compared with CA-RSRP (Coupled Access based on Reference Signal Received Power), DUDA-RSRP (Downlink and Uplink Decoupled Access based on Reference Signal Received Power), and QMIX-DUDA (QMIX-based Downlink and Uplink Decoupled Access). The comparison results are as follows: Figure 5 and Figure 6 As shown, where, Figure 5 This is a comparison of the switching stability metrics of different algorithms. Figure 5 (a) in the figure is a comparison chart of the number of switching operations for different algorithms. Figure 5 (b) in the figure is a comparison chart of the failure switching ratios of different algorithms. Figure 5 (c) in the figure is a comparison chart of the ping-pong switching ratios of different algorithms; Figure 6 This is a comparison of the overall performance metrics of different algorithms. Figure 6 (a) in the figure is a comparison of the sum rate of different algorithms. Figure 6 (b) in the figure is a comparison chart of the fairness of different algorithms. Figure 6 (c) in the figure is a comparison of the decoupling ratios of different algorithms.

[0040] Step S30: Calculate the local action value of the uplink local observation and the downlink local observation to obtain the target uplink local action value and the target downlink local action value, and perform joint processing on the target uplink local action value and the target downlink local action value to obtain the target joint action value.

[0041] Specifically, after inputting the uplink local observation and the downlink local observation into the trained uplink-downlink decoupling handover decision network, the uplink-downlink decoupling handover decision network calculates the local action value of the uplink local observation and the downlink local observation to obtain the target uplink local action value and the target downlink local action value, and performs joint processing on the target uplink local action value and the target downlink local action value to obtain the target joint action value.

[0042] Step S40: Map the uplink candidate actions corresponding to the uplink local observation and the downlink candidate actions corresponding to the downlink local observation to the target interval to obtain a security score result representing the security confidence of the candidate actions. Then, fuse the joint action value and the security score result to obtain the target action evaluation value. Finally, perform network switching based on the target switching action of the target action evaluation value to obtain the network intelligent switching result.

[0043] Specifically, the uplink candidate actions corresponding to the uplink local observations and the downlink candidate actions corresponding to the downlink local observations are mapped to the target interval to obtain a security score result representing the security confidence of the candidate actions. The target interval is the [0, 1] interval, and the security score result includes the target uplink security score and the target downlink security score. The joint action value, the target uplink security score, and the target downlink security score are fused to obtain the target action evaluation value. Network switching is performed based on the target switching action of the target action evaluation value to obtain the network intelligent switching result.

[0044] The beneficial technical effects of the present invention are as follows: (1) This invention allows users to select serving base stations in the uplink and downlink directions respectively, which can better adapt to uplink and downlink asymmetry, base station type differences and user mobility changes, thereby improving problems such as limited uplink coverage, uneven load distribution and reduced link quality for edge users; (2) The present invention adopts a multi-agent reinforcement learning framework based on GRU-QMIX for handover decision-making. GRU is used to extract the temporal features of link status and user movement, and QMIX is used to characterize the cooperative relationship of multiple base stations, so that each base station can learn the global cooperative strategy during the training phase and complete the distributed decision-making based on local observations during the execution phase, thereby improving the adaptability and deployability of the handover strategy. (3) The present invention introduces a subtractive clustering reward smoothing mechanism, which generates a smooth reward based on the distribution characteristics of key indicators such as system rate, link quality, fairness and handover stability, reducing the interference of instantaneous channel fluctuations, abnormal samples and occasional handover events on the training process, thereby improving the model training stability and policy convergence quality, without significantly increasing online execution overhead. (4) The present invention introduces a spatiotemporal attention residual action screening network to perform risk perception assessment on candidate handover actions and integrates action value with security score to generate final decision. This can suppress high-risk behaviors such as handover failure, ping-pong handover, link quality degradation or load deterioration, thereby improving handover reliability and service stability while improving system transmission rate and user fairness.

[0045] Furthermore, such as Figure 7 As shown, based on the above-described intelligent network handover method for dynamic wireless environments, the present invention also provides a corresponding intelligent network handover system for dynamic wireless environments, wherein the intelligent network handover system for dynamic wireless environments includes: The target base station confirmation module 51 is used to obtain the user's network status, determine the target base station based on the network status, obtain the uplink target base station and the downlink target base station, and confirm the handover behavior based on the uplink target base station and the downlink target base station to obtain the handover confirmation result. The local observation acquisition module 52 is used to acquire the uplink local observation of the uplink target base station and the downlink local observation of the downlink target base station according to the handover confirmation result if the handover confirmation result indicates that a handover action exists, and input the uplink local observation and the downlink local observation into the trained uplink-downlink decoupled handover decision network. Action value calculation module 53 is used to calculate the local action value of the uplink local observation and the downlink local observation to obtain the target uplink local action value and the target downlink local action value, and to perform joint processing on the target uplink local action value and the target downlink local action value to obtain the target joint action value; The network intelligent handover module 54 is used to map the uplink candidate actions corresponding to the uplink local observation and the downlink candidate actions corresponding to the downlink local observation to the target interval, obtain a security score result representing the security confidence of the candidate actions, fuse the joint action value and the security score result to obtain the target action evaluation value, and perform network handover according to the target handover action of the target action evaluation value to obtain the network intelligent handover result.

[0046] Furthermore, such as Figure 8 As shown, based on the above-described intelligent network handover method for dynamic wireless environments, the present invention also provides a terminal, which includes a processor 10, a memory 20, and a display 30. Figure 8 Only some of the terminal components are shown; however, it should be understood that it is not required to implement all of the components shown, and more or fewer components may be implemented instead.

[0047] In some embodiments, the memory 20 may be an internal storage unit of the terminal, such as a hard disk or memory. In other embodiments, the memory 20 may be an external storage device of the terminal, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc. Further, the memory 20 may include both internal and external storage devices. The memory 20 is used to store application software and various types of data installed on the terminal, such as program code installed on the terminal. The memory 20 can also be used to temporarily store data that has been output or will be output. In one embodiment, the memory 20 stores a network intelligent handover program 40 for dynamic wireless environments, which can be executed by the processor 10 to implement the network intelligent handover method for dynamic wireless environments in this application.

[0048] In some embodiments, the processor 10 may be a central processing unit (CPU), a microprocessor, or other data processing chip, used to run program code stored in the memory 20 or process data, such as executing the network intelligent handover method for dynamic wireless environments.

[0049] In some embodiments, the display 30 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen. The display 30 is used to display information on the terminal and to display a visual user interface.

[0050] In one embodiment, when the processor 10 executes the network intelligent handover program 40 for dynamic wireless environments stored in the memory 20, the following steps are performed: The user's network status is obtained, the target base station is determined based on the network status, the uplink target base station and the downlink target base station are obtained, and the handover behavior is confirmed based on the uplink target base station and the downlink target base station to obtain the handover confirmation result; If the handover confirmation result indicates that a handover action exists, then the uplink local observation of the uplink target base station and the downlink local observation of the downlink target base station are obtained based on the handover confirmation result, and the uplink local observation and the downlink local observation are input into the trained uplink-downlink decoupled handover decision network. Local action values ​​are calculated for the uplink local observations and the downlink local observations to obtain the target uplink local action value and the target downlink local action value. The target uplink local action value and the target downlink local action value are then jointly processed to obtain the target joint action value. The uplink candidate actions corresponding to the uplink local observations and the downlink candidate actions corresponding to the downlink local observations are mapped to the target interval to obtain a security score result representing the security confidence of the candidate actions. The joint action value and the security score result are fused to obtain the target action evaluation value. The network switching is performed according to the target switching action of the target action evaluation value to obtain the network intelligent switching result.

[0051] The process of obtaining the user's network status, determining the target base station based on the network status to obtain the uplink target base station and the downlink target base station, and confirming the handover behavior based on the uplink target base station and the downlink target base station to obtain the handover confirmation result specifically includes: The network status of the user's location is obtained, and uplink and downlink base stations are associated with the network status based on the base station agent of the area to obtain the base station association result. Based on the base station association result, the base station is determined to obtain the uplink target base station and the downlink target base station. The number of handovers of the uplink target base station and the downlink target base station is statistically processed to obtain handover statistics results. Based on the handover statistics results, the handover behavior of the uplink target base station and the downlink target base station is confirmed to obtain handover confirmation results.

[0052] The training process of the uplink / downlink decoupling handover decision network specifically includes: Obtain a network training set and construct an uplink / downlink decoupling decision training network. The network training set includes multiple sets of historical uplink local observations of historical uplink base stations and historical downlink local observations of historical downlink base stations. The uplink / downlink decoupling decision training network includes an uplink / downlink decoupling association network and a spatiotemporal attention residual action filtering network. A set of historical uplink local observations and historical downlink local observations from the network training set are input into the uplink-downlink decoupling decision training network. The uplink-downlink decoupling correlation network encodes the historical uplink local observations and historical downlink local observations respectively to obtain the uplink temporal hidden state and the downlink temporal hidden state. Action values ​​are calculated for the uplink and downlink hidden timing states to obtain uplink local action values ​​and downlink local action values. The uplink local action values ​​and downlink local action values ​​are then combined to obtain a combined action value. Set the original reward, and smooth the original reward according to the subtractive clustering reward smoothing mechanism to obtain the smoothed reward; The spatiotemporal attention residual action filtering network is used to construct features for the historical uplink candidate actions corresponding to the historical uplink local observations and the historical downlink candidate actions corresponding to the historical downlink local observations, so as to obtain uplink risk assessment features and downlink risk assessment features. The uplink risk assessment features and downlink risk assessment features are then mapped to the target interval to obtain uplink security scores and downlink security scores that characterize the security confidence of candidate actions. The combined action value, the uplink security score, and the downlink security score are fused together to obtain a handover action evaluation value. A network intelligent handover action is obtained based on the handover action evaluation value. The network intelligent handover action is executed to obtain an action execution result. The uplink and downlink association of the user is updated based on the execution result. The parameters of the uplink-downlink decoupling decision training network are updated according to the smoothing reward, and the next set of historical uplink local observations and historical downlink local observations are input into the uplink-downlink decoupling decision training network until the training status of the uplink-downlink decoupling decision training network meets the preset conditions, thus obtaining the trained uplink-downlink decoupling decision network.

[0053] Specifically, setting the original reward and smoothing it using a subtractive clustering reward smoothing mechanism to obtain a smoothed reward includes: Set the initial reward and obtain key communication indicators, and construct a communication status sample based on the key communication indicators; Obtain a sample set, calculate the density potential value of each sample point in the sample set to obtain the calculation result, extract the maximum density potential value in the calculation result, and reduce the samples in the sample set according to the target sample corresponding to the maximum density potential value to obtain a cluster center set. The distance between the communication state sample and each cluster center in the cluster center set is calculated to obtain the distance calculation result. The membership weight is calculated based on the distance calculation result, and the original reward is smoothed based on the membership weight to obtain the smoothed reward.

[0054] Specifically, setting the original reward means: ; The calculation of the density potential value for each sample point in the sample set is specifically as follows: ; The step of reducing the number of samples in the sample set based on the target sample corresponding to the maximum density potential value specifically involves: ; in, For the original reward, The normalized system transmission rate, The normalized average link quality metric. As a fairness metric used to evaluate load balancing among base stations, This represents the normalized number of handover failures. This represents the normalized number of ping-pong switches. , , , and All are reward weighting coefficients. For the sample density potential, For the first One sample, For the first One sample, To influence the radius of clustering, For the first The residual density potential of each sample after density reduction. The density potential value corresponding to the cluster center. As cluster center, The radius is the density reduction radius.

[0055] The process of mapping the uplink risk assessment features and the downlink risk assessment features to a target interval to obtain uplink security scores and downlink security scores that characterize the security confidence of candidate actions further includes: The historical uplink candidate actions and the historical downlink candidate actions are executed to obtain action execution results, and risk labels are constructed based on the execution results; The training loss of the spatiotemporal attention residual action screening network is calculated based on the risk label to obtain the loss value, and the spatiotemporal attention residual action screening network is optimized based on the loss value.

[0056] The step of updating the parameters of the uplink / downlink decoupling decision training network based on the smoothing reward specifically includes: Obtain the target batch of samples, and calculate the temporal difference target of the target batch of samples based on the smoothing reward; The network training loss is calculated based on the temporal difference objective to obtain the target loss value, and the parameters of the uplink / downlink decoupling decision training network are updated based on the target loss value.

[0057] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a network intelligent handover program for dynamic wireless environments, and the network intelligent handover program for dynamic wireless environments, when executed by a processor, implements the steps of the network intelligent handover method for dynamic wireless environments as described above.

[0058] In summary, this invention provides a network intelligent handover method, system, terminal, and storage medium for dynamic wireless environments. The method includes: acquiring a user's network state; determining target base stations based on the network state to obtain uplink target base stations and downlink target base stations; confirming handover behavior based on the uplink target base stations and downlink target base stations to obtain a handover confirmation result; if the handover confirmation result indicates that handover behavior exists, acquiring uplink local observations of the uplink target base station and downlink local observations of the downlink target base station based on the handover confirmation result, and inputting the uplink local observations and downlink local observations into a trained uplink-downlink decoupled handover decision network; and for the... Local action values ​​are calculated using uplink and downlink local observations to obtain target uplink and downlink local action values. These values ​​are then jointly processed to obtain a target joint action value. Uplink candidate actions corresponding to the uplink local observations and downlink candidate actions corresponding to the downlink local observations are mapped to target intervals to obtain a security score representing the security confidence of the candidate actions. The joint action value and the security score are then fused to obtain a target action evaluation value. Network switching is then performed based on the target switching action according to the target action evaluation value, resulting in a network intelligent switching result. This invention can simultaneously achieve adaptive smoothing of reward feedback and risk-aware screening of candidate switching actions, improving the success rate and reliability of network intelligent switching.

[0059] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0060] Of course, those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware (such as a processor, controller, etc.). The program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The computer-readable storage medium can be a memory, magnetic disk, optical disk, etc.

[0061] It should be understood that the application of the present invention is not limited to the examples above. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A network intelligent handover method for dynamic wireless environments, characterized in that, The network intelligent handover method for dynamic wireless environments includes: The user's network status is obtained, the target base station is determined based on the network status, the uplink target base station and the downlink target base station are obtained, and the handover behavior is confirmed based on the uplink target base station and the downlink target base station to obtain the handover confirmation result; If the handover confirmation result indicates that a handover action exists, then the uplink local observation of the uplink target base station and the downlink local observation of the downlink target base station are obtained based on the handover confirmation result, and the uplink local observation and the downlink local observation are input into the trained uplink-downlink decoupled handover decision network. Local action values ​​are calculated for the uplink local observations and the downlink local observations to obtain the target uplink local action value and the target downlink local action value. The target uplink local action value and the target downlink local action value are then jointly processed to obtain the target joint action value. The uplink candidate actions corresponding to the uplink local observations and the downlink candidate actions corresponding to the downlink local observations are mapped to the target interval to obtain a security score result representing the security confidence of the candidate actions. The joint action value and the security score result are fused to obtain the target action evaluation value. The network switching is performed according to the target switching action of the target action evaluation value to obtain the network intelligent switching result.

2. The network intelligent handover method for dynamic wireless environments according to claim 1, characterized in that, The process of obtaining the user's network status, determining the target base station based on the network status to obtain the uplink target base station and the downlink target base station, and confirming the handover behavior based on the uplink target base station and the downlink target base station to obtain the handover confirmation result specifically includes: The network status of the user's location is obtained, and uplink and downlink base stations are associated with the network status based on the base station agent of the area to obtain the base station association result. Based on the base station association result, the base station is determined to obtain the uplink target base station and the downlink target base station. The number of handovers of the uplink target base station and the downlink target base station is statistically processed to obtain handover statistics results. Based on the handover statistics results, the handover behavior of the uplink target base station and the downlink target base station is confirmed to obtain handover confirmation results.

3. The network intelligent handover method for dynamic wireless environments according to claim 1, characterized in that, The training process of the uplink / downlink decoupling handover decision network specifically includes: Obtain a network training set and construct an uplink / downlink decoupling decision training network. The network training set includes multiple sets of historical uplink local observations of historical uplink base stations and historical downlink local observations of historical downlink base stations. The uplink / downlink decoupling decision training network includes an uplink / downlink decoupling association network and a spatiotemporal attention residual action filtering network. A set of historical uplink local observations and historical downlink local observations from the network training set are input into the uplink-downlink decoupling decision training network. The uplink-downlink decoupling correlation network encodes the historical uplink local observations and historical downlink local observations respectively to obtain the uplink temporal hidden state and the downlink temporal hidden state. Action values ​​are calculated for the uplink and downlink hidden timing states to obtain uplink local action values ​​and downlink local action values. The uplink local action values ​​and downlink local action values ​​are then combined to obtain a combined action value. Set the original reward, and smooth the original reward according to the subtractive clustering reward smoothing mechanism to obtain the smoothed reward; The spatiotemporal attention residual action filtering network is used to construct features for the historical uplink candidate actions corresponding to the historical uplink local observations and the historical downlink candidate actions corresponding to the historical downlink local observations, so as to obtain uplink risk assessment features and downlink risk assessment features. The uplink risk assessment features and downlink risk assessment features are then mapped to the target interval to obtain uplink security scores and downlink security scores that characterize the security confidence of candidate actions. The combined action value, the uplink security score, and the downlink security score are fused together to obtain a handover action evaluation value. A network intelligent handover action is obtained based on the handover action evaluation value. The network intelligent handover action is executed to obtain an action execution result. The uplink and downlink association of the user is updated based on the execution result. The parameters of the uplink-downlink decoupling decision training network are updated according to the smoothing reward, and the next set of historical uplink local observations and historical downlink local observations are input into the uplink-downlink decoupling decision training network until the training status of the uplink-downlink decoupling decision training network meets the preset conditions, thus obtaining the trained uplink-downlink decoupling decision network.

4. The network intelligent handover method for dynamic wireless environments according to claim 3, characterized in that, The process of setting the original reward and smoothing it using a subtractive clustering reward smoothing mechanism to obtain a smoothed reward specifically includes: Set the initial reward and obtain key communication indicators, and construct a communication status sample based on the key communication indicators; Obtain a sample set, calculate the density potential value of each sample point in the sample set to obtain the calculation result, extract the maximum density potential value in the calculation result, and reduce the samples in the sample set according to the target sample corresponding to the maximum density potential value to obtain a cluster center set. The distance between the communication state sample and each cluster center in the cluster center set is calculated to obtain the distance calculation result. The membership weight is calculated based on the distance calculation result, and the original reward is smoothed based on the membership weight to obtain the smoothed reward.

5. The network intelligent handover method for dynamic wireless environments according to claim 4, characterized in that, The setting of the original reward is as follows: ； The calculation of the density potential value for each sample point in the sample set is specifically as follows: ； The step of reducing the number of samples in the sample set based on the target sample corresponding to the maximum density potential value specifically involves: ； in, For the original reward, The normalized system transmission rate, The normalized average link quality metric. As a fairness metric used to evaluate load balancing among base stations, This represents the normalized number of handover failures. This represents the normalized number of ping-pong switches. , , , and All of these are reward weighting coefficients. For the sample density potential, For the first One sample, For the first One sample, To influence the radius of clustering, For the first The residual density potential of each sample after density reduction. The density potential value corresponding to the cluster center. As cluster center, The radius is the density reduction radius.

6. The network intelligent handover method for dynamic wireless environments according to claim 3, characterized in that, The process of mapping the uplink risk assessment features and the downlink risk assessment features to a target interval to obtain an uplink security score and a downlink security score representing the security confidence of candidate actions further includes: The historical uplink candidate actions and the historical downlink candidate actions are processed to obtain action execution results, and risk labels are constructed based on the execution results; The training loss of the spatiotemporal attention residual action screening network is calculated based on the risk label to obtain the loss value, and the spatiotemporal attention residual action screening network is optimized based on the loss value.

7. The network intelligent handover method for dynamic wireless environments according to claim 3, characterized in that, The step of updating the parameters of the uplink / downlink decoupling decision training network based on the smoothing reward specifically includes: Obtain the target batch of samples, and calculate the temporal difference target of the target batch of samples based on the smoothing reward; The network training loss is calculated based on the temporal difference objective to obtain the target loss value, and the parameters of the uplink / downlink decoupling decision training network are updated based on the target loss value.

8. A network intelligent handover system for dynamic wireless environments, characterized in that, The network intelligent handover system for dynamic wireless environments includes: The target base station confirmation module is used to obtain the user's network status, determine the target base station based on the network status, obtain the uplink target base station and the downlink target base station, and confirm the handover behavior based on the uplink target base station and the downlink target base station to obtain the handover confirmation result. The local observation acquisition module is used to acquire uplink local observations of the uplink target base station and downlink local observations of the downlink target base station based on the handover confirmation result if the handover confirmation result indicates that a handover action exists, and input the uplink local observations and the downlink local observations into the trained uplink-downlink decoupled handover decision network. The action value calculation module is used to calculate the local action value of the uplink local observation and the downlink local observation to obtain the target uplink local action value and the target downlink local action value, and to perform joint processing on the target uplink local action value and the target downlink local action value to obtain the target joint action value; The network intelligent handover module is used to map the uplink candidate actions corresponding to the uplink local observation and the downlink candidate actions corresponding to the downlink local observation to the target interval, obtain a security score result representing the security confidence of the candidate actions, fuse the joint action value and the security score result to obtain the target action evaluation value, and perform network handover according to the target handover action of the target action evaluation value to obtain the network intelligent handover result.

9. A terminal, characterized in that, The terminal includes a memory, a processor, and a program stored in the memory and executable on the processor. When the program is executed by the processor, it implements the steps of the network intelligent handover method for dynamic wireless environments as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program thereon, and the computer-readable storage medium stores a network intelligent handover program for dynamic wireless environments. When the network intelligent handover program for dynamic wireless environments is executed by a processor, it implements the steps of the network intelligent handover method for dynamic wireless environments as described in any one of claims 1-7.

Images