Unlicensed frequency band conflict modeling and backoff optimization method and device based on graph neural network

Abstract

The embodiment of the invention discloses an unlicensed frequency band conflict modeling and backoff optimization method and device based on a graph neural network, and aims to solve the conflict problem when multiple devices share a frequency spectrum in an unlicensed frequency band. The method comprises the steps of collecting and preprocessing multi-dimensional information of wireless equipment, constructing a time-space associated conflict graph, predicting a conflict probability by using a graph attention network, optimizing a backoff strategy through multi-agent reinforcement learning, and feeding back and updating a model in real time. The device comprises a data acquisition unit, a data processing unit, a graph construction unit, a conflict modeling unit, a strategy optimization unit and a feedback optimization unit. Through integration of the graph neural network and multi-agent reinforcement learning, precise modeling of conflicts and adaptive optimization of a backoff strategy are realized, the spectrum utilization rate and the system throughput on an unlicensed frequency band are significantly improved, the probability of conflicts between devices is reduced, and the method is suitable for high-density wireless communication scenes such as the Internet of Things and smart home.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of wireless communication, in particular to a graph neural network-based unlicensed frequency band conflict modeling and backoff optimization method and device, which is used to solve the conflict problem when multiple devices share the spectrum in the unlicensed frequency band, and improve the data transmission efficiency and system performance. BACKGROUND

[0002] With the rapid development of wireless communication technology, emerging applications such as Internet of Things, smart home, and wireless sensor networks are constantly emerging, and the demand for spectrum resources is increasing dramatically. Due to the limited number and strict allocation of licensed spectrum resources, it is difficult to meet the growing communication demand, so the unlicensed frequency band becomes a key resource for many wireless devices to share. However, in the unlicensed frequency band environment, a large number of wireless devices access at the same time, which is easy to cause channel conflict and interference, and seriously affects the efficiency and stability of data transmission.

[0003] In the multi-user sharing scenario of unlicensed frequency bands (such as 2.4GHz / 5GHz Wi-Fi, WAPI, etc.), the binary exponential backoff (BEB) mechanism as the core technology of the carrier sense multiple access / collision avoidance (CSMA / CA) protocol, through dynamic adjustment of the backoff window to alleviate channel competition conflict, is the basis for ensuring that devices can access the unlicensed frequency band fairly and friendly. For example, the IEEE 802.11 standard uses the BEB mechanism to realize channel access control of devices in the uncoordinated contention period, and its core idea is to reduce the retransmission conflict probability by exponentially increasing the backoff window, which shows good robustness in traditional low-load scenarios. However, with the rapid increase in the density of Internet of Things (IoT) devices (such as hundreds of sensors simultaneously accessing in a smart factory), the independent decision-making characteristics of the BEB mechanism lead to the problem of "backoff storm" frequently, which reduces the spectrum utilization.

[0004] Currently, for conflict resolution and backoff optimization in the unlicensed frequency band, traditional methods mostly use fixed backoff algorithms, such as the binary exponential backoff (BEB) algorithm. Taking a practical Internet of Things scenario as an example, when there are 20 devices in the network competing for the channel at the same time, using the BEB algorithm, in the case of frequent device access, the average conflict number is as high as 15 times per 100 time slots, resulting in a channel utilization rate of only about 40%. This is because although the BEB algorithm is simple and easy to implement, in a complex and variable wireless environment, its backoff window adjustment only depends on the number of conflicts, and cannot effectively adapt to changes in network topology and dynamic device access, resulting in a lack of flexibility in backoff strategies, which easily causes waste of channel resources and increases transmission delay.

[0005] Some machine learning-based methods have also been applied in this field, such as traditional neural network-based backoff strategy optimization methods. However, such methods usually treat wireless devices as independent individuals and fail to fully consider the spatial interference relationship and historical conflict patterns between devices. In a simulation experiment of an unlicensed frequency band network containing multiple cells, using a traditional neural network method, due to the neglect of the signal interference between devices in adjacent cells, the conflict probability of the edge devices of the cell is higher than that of the center devices, which reduces the overall network throughput. This shows that traditional machine learning methods are difficult to fully model complex conflict scenarios and cannot achieve efficient backoff optimization, which cannot meet the requirements of modern wireless communication for high reliability and high efficiency. Therefore, there is an urgent need for a method and device that can accurately model conflicts and achieve intelligent backoff optimization. SUMMARY

[0006] The purpose of the present application is to provide a graph neural network-based unlicensed frequency band conflict modeling and backoff optimization method and device, which realizes accurate modeling of device conflicts in unlicensed frequency bands and adaptive optimization of backoff strategies by constructing a spatio-temporally correlated conflict graph structure, combining graph attention mechanisms and multi-agent reinforcement learning, and significantly improving spectrum utilization and system throughput.

[0007] The first aspect of the embodiment of the present application provides a graph neural network-based unlicensed frequency band conflict modeling and backoff optimization method, comprising:

[0008] 1. The data acquisition module acquires multi-dimensional information of wireless devices, including device unique identifier, geographic position coordinates, transmission power, transmission rate, current working channel number, communication state of each time slot and time stamp;

[0009] 2. The data processing module pre-processes the collected data, including rejecting abnormal data, normalizing continuous data, encoding discrete data, and dividing data according to time window, extracting time and space dimension features, and fusing to form device feature vectors containing spatio-temporal correlation;

[0010] 3. The construction module defines each wireless device as a graph node, generates edges according to the distance between devices, historical conflicts and channel sharing, and calculates edge weights to construct a spatio-temporally correlated conflict graph;

[0011] 4. The conflict graph is processed based on the graph attention network, and the node features containing the complex correlation relationship of the devices are output through multi-layer calculation for predicting the conflict probability of the devices;

[0012] 5. Each wireless device is regarded as an agent, and a multi-agent deep deterministic policy gradient algorithm is used for backoff strategy learning, and each agent executes the optimal backoff action according to the current state;

[0013] 6. Real-time monitoring of device communication status and system performance indicators, if performance is not as expected, retrigger multi-agent reinforcement learning to update backoff strategy, update conflict graph structure and graph neural network model parameters periodically or when network topology changes significantly.

[0014] With reference to the first aspect, in a first possible implementation manner of the first aspect, the data collection module acquires multi-dimensional information of the wireless device, including: the data collection module periodically collects device basic attribute information through a built-in communication module of the wireless device, and records device communication state information in real time.

[0015] With reference to the first aspect, in a second possible implementation manner of the first aspect, the data processing module normalizes the continuous data, specifically including: the data processing module uses a maximum-minimum normalization method to map the continuous data to the interval [0, 1].

[0016] With reference to the first aspect, in a third possible implementation manner of the first aspect, the construction module generates edges according to the distance between devices, historical conflicts and channel sharing conditions, including: if the distance between two devices is less than a preset interference threshold, or multiple conflicts have occurred in continuous time windows, or the same working channel is shared in the same time window and there is a data transmission behavior, an edge is added between the corresponding nodes.

[0017] With reference to the third possible implementation manner of the first aspect, in a fourth possible implementation manner of the first aspect, the calculation of the edge weight also considers the time synchronization degree of the device state change, which is reflected through a time correlation degree parameter.

[0018] With reference to the first aspect, the first possible implementation manner of the first aspect, the second possible implementation manner of the first aspect, the third possible implementation manner of the first aspect or the fourth possible implementation manner of the first aspect, in a fifth possible implementation manner of the first aspect, the graph attention network is provided with multiple layers of attention mechanism layers and multiple attention heads, and the importance degree between nodes is adaptively learned through attention calculation.

[0019] With reference to the fifth possible implementation manner of the first aspect, in a sixth possible implementation manner of the first aspect, the conflict graph is processed based on the graph attention network, including: performing linear transformation on node features, calculating node attention coefficients and normalizing, and weighting and aggregating neighbor node features to update current node features.

[0020] With reference to the first aspect, in a seventh possible implementation manner of the first aspect, in the multi-agent deep deterministic policy gradient algorithm, each agent maintains a policy network and a value network, and updates network parameters through sampling of state-action sequences of the agent.

[0021] With reference to the eighth possible implementation manner of the first aspect, in a ninth possible implementation manner of the first aspect, if the collision rate exceeds a preset threshold, the multi-agent reinforcement learning is retriggered to update the backoff strategy.

[0022] With reference to the eighth possible implementation manner of the first aspect, in a ninth possible implementation manner of the first aspect, if the collision rate exceeds a preset threshold, the multi-agent reinforcement learning is retriggered to update the backoff strategy.

[0023] The second aspect of the embodiment of the present application provides another method for modeling and optimizing backoff of unlicensed frequency band collision based on a graph neural network, applied to a wireless device end, comprising:

[0024] 1. The wireless device collects its own information and sends it to a data processing center.

[0025] 2. The wireless device receives a backoff strategy adjustment instruction sent by the data processing center.

[0026] 3. The wireless device adjusts its backoff window size, switches a working channel, or changes a transmission power according to the backoff strategy adjustment instruction.

[0027] With reference to the second aspect, in a first possible implementation manner of the second aspect, the wireless device collects its own information, including a device unique identifier, a geographic position coordinate, a transmission power, a current working channel number, and communication state information.

[0028] With reference to the second aspect, in a second possible implementation manner of the second aspect, the backoff strategy adjustment instruction includes target backoff window size, target working channel number, and the like.

[0029] The third aspect of the embodiment of the present application provides a device for modeling and optimizing backoff of unlicensed frequency band collision based on a graph neural network, comprising:

[0030] 1. A data collection unit configured to acquire multi-dimensional information of a wireless device.

[0031] 2. A data processing unit configured to preprocess collected data and extract spatiotemporal features.

[0032] 3. A graph construction unit configured to construct a conflict graph associated with space and time.

[0033] 4. A conflict modeling unit configured to process the conflict graph based on a graph attention network.

[0034] 5. A strategy optimization unit configured to optimize a backoff strategy by using multi-agent reinforcement learning.

[0035] 6. A feedback optimization unit configured to monitor system performance in real time and update strategy and model parameters.

[0036] In a possible design, a structure of the apparatus for modeling and optimizing backoff of unlicensed band conflict based on a graph neural network includes a processor and a communication interface, where the processor is configured to execute the method for modeling and optimizing backoff of unlicensed band conflict based on a graph neural network provided in the first aspect of the present application. Optionally, the apparatus can further include a memory configured to store application code for supporting the apparatus to execute the method, and the processor is configured to execute the application stored in the memory.

[0037] The fourth aspect of the embodiment of the present application provides a wireless device, including:

[0038] an information collection unit configured to collect information of the wireless device itself;

[0039] an instruction receiving unit configured to receive a backoff policy adjustment instruction;

[0040] a policy executing unit configured to execute the backoff policy according to the backoff policy adjustment instruction.

[0041] In a possible design, a structure of the wireless device includes a processor and a communication module, where the processor is configured to execute the method for modeling and optimizing backoff of unlicensed band conflict based on a graph neural network provided in the second aspect of the present application. Optionally, the wireless device can further include a memory configured to store application code for supporting the wireless device to execute the method, and the processor is configured to execute the application stored in the memory.

[0042] The fifth aspect of the embodiment of the present application provides a computer storage medium, which stores program code, and the program code is executed by a computing device to execute the method for modeling and optimizing backoff of unlicensed band conflict based on a graph neural network provided in the first aspect.

[0043] The sixth aspect of the embodiment of the present application provides another computer storage medium, which stores program code, and the program code is executed by a computing device to execute the method for modeling and optimizing backoff of unlicensed band conflict based on a graph neural network provided in the second aspect.

[0044] In the embodiment of the present application, the names of the data collection module, the data processing module and the like do not constitute a limitation on the embodiment of the present application, and in actual implementation, these modules can appear in other names. As long as the functions of each module are similar to those of the present application, they belong to the scope of the claims of the present application and equivalent technologies.

[0045] In the embodiment of the present application, by constructing a conflict graph structure with space-time correlation, combining a graph attention mechanism and multi-agent reinforcement learning, the spatial interference relationship between devices, historical conflict patterns and time synchronization can be fully considered, the problem of "backoff storm" easily caused by the traditional BEB mechanism in high-density scenarios is solved, precise modeling of device conflicts in the unlicensed frequency band and adaptive optimization of backoff strategies are realized, compared with traditional methods, the spectrum utilization and system throughput are significantly improved, and the conflict probability between devices in the unlicensed frequency band is reduced. At the same time, through the real-time monitoring and dynamic updating mechanism, the system can adapt to network topology changes, and the stability and efficiency of the wireless communication system in the unlicensed frequency band environment are enhanced, which is suitable for emerging high-density wireless communication scenarios such as Internet of Things and smart home. BRIEF DESCRIPTION OF DRAWINGS

[0046] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0047] Figure 1 is a flow chart of unlicensed frequency band conflict modeling and backoff optimization based on a graph neural network provided by the embodiment of the present application;

[0048] Figure 2 is a space-time correlation conflict graph construction process provided by the embodiment of the present application;

[0049] Figure 3 is a conflict modeling process based on a graph attention network (GAT) provided by the embodiment of the present application;

[0050] Figure 4 is a backoff strategy optimization process based on multi-agent reinforcement learning provided by the embodiment of the present application;

[0051] Figure 5 is a system optimization process based on dynamic feedback provided by the embodiment of the present application;

[0052] Figure 6 is an unlicensed frequency band conflict modeling and backoff optimization device based on a graph neural network provided by the embodiment of the present application. DETAILED DESCRIPTION

[0053] With reference to the drawings and embodiments of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative efforts belong to the scope of the present application.

[0054] It should be understood that the terms "comprise" and "include" as used in the specification and the appended claims indicate the presence of the described features, integers, steps, operations, elements, and / or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0055] It should also be understood that the terms used in the present application specification are only for the purpose of describing specific embodiments and are not intended to limit the present application. As used in the present application specification and the appended claims, the singular forms "a", "an" and "the" are intended to include the plural forms unless the context clearly indicates otherwise.

[0056] It should be further understood that the term "and / or" as used in the present application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations thereof.

[0057] As used in the present application specification and the appended claims, the term "if" can be interpreted as "when" or "upon" or "in response to determining" or "in response to detecting" depending on the context. Similarly, the phrases "if it is determined" or "if [a described condition or event] is detected" can be interpreted to mean "upon determining" or "in response to determining" or "upon detecting [a described condition or event]" or "in response to detecting [a described condition or event]" depending on the context.

[0058] To facilitate understanding of the present application, the following first introduces the graph neural network-based unlicensed frequency band conflict modeling and backoff optimization method provided by the embodiments of the present application, as shown in Figure 1 The specific implementation process of the graph neural network-based unlicensed frequency band conflict modeling and backoff optimization method is as follows:

[0059] Step 1: Data acquisition and spatio-temporal feature extraction: Use the wireless device communication module to collect multi-dimensional data such as device basic attributes, communication state and time information. Divide the data according to the time window, process the time and space dimension features respectively, and finally fuse to form a device feature vector containing spatio-temporal correlation, providing data basis for subsequent modeling.

[0060] Step 2: Construction of spatio-temporal conflict graph: Each wireless device is defined as a graph node, and the node features integrate spatio-temporal information. Edges are generated based on device distance, historical conflicts, and channel sharing, and edge weights are determined by combining interference strength, conflict frequency, and time correlation to construct an undirected weighted graph. At the same time, the graph structure is updated in real time as the device and channel states change.

[0061] Step 3: Conflict modeling based on graph attention mechanism: A multi-layer graph attention network (GAT) architecture is built to update the current node features by performing linear transformation on node features, calculating attention coefficients and normalization, and weighted aggregation of neighbor node features. The output includes device complex correlation relationships, which are used to predict device conflict situations.

[0062] Step 4: Multi-agent reinforcement learning backoff strategy optimization: Each wireless device is considered as an agent, and the agent state, environment, action, and reward are defined. The multi-agent deep deterministic policy gradient (MADDPG) algorithm is used, and each agent maintains a policy network and a value network. By sampling and updating network parameters, the agents learn a cooperative backoff strategy, and execute the optimal action according to the current state.

[0063] Step 5: Real-time feedback and continuous optimization: Real-time monitoring of device communication state and system performance indicators is performed. If the performance does not meet expectations, the multi-agent reinforcement learning is retriggered to update the backoff strategy. The conflict graph structure and graph neural network model parameters are updated periodically or when the network topology changes significantly, ensuring continuous and efficient system operation.

[0064] After data collection, the physical devices and communication states need to be abstracted into a graph structure to provide structured input for subsequent conflict modeling. The key step is to construct a conflict graph that reflects the spatio-temporal association of devices, as shown in Example 2. Example 2 provides a flowchart for constructing a spatio-temporal conflict graph. This example focuses on node definition, edge generation, and graph structure updating. By integrating multi-dimensional device information, a dynamic conflict graph reflecting device relationships is constructed, providing a foundation for subsequent processing. The detailed steps and logic are as follows.

[0065] Step 1: Data collection and preprocessing

[0066] Collecting multi-dimensional device information: Through the built-in communication module, periodically acquire device ID, coordinates, channel number, transmission power, and other basic attributes. Real-time record of communication state (success / conflict), RSSI, and timestamp of each time slot.

[0067] Data preprocessing: Remove abnormal data, normalize continuous data (such as coordinates, power) to [0,1] using max-min normalization, one-hot encode discrete data (such as channel number), encode conflict state as 0 / 1, and divide data by time window (such as 10 seconds).

[0068] Step 2: Node Definition and Feature Fusion

[0069] Define graph nodes: Each device corresponds to a node, and the node features fuse space-time information to generate a vector such as [normalized coordinates, channel one-hot encoding, conflict frequency in the time window, successful transmission frequency].

[0070] Step 3: Edge Generation and Condition Judgment

[0071] Connection conditions (satisfy any one to add edges):

[0072] Spatial interference: The distance between device i (coordinates x i ,y i ) and device j (coordinates x j ,y j ) is less than a preset threshold (such as 50 meters), calculated by the Euclidean distance formula: .

[0073] Historical conflict: Conflicts ≥ 2 times in the last 3 time windows (the number of conflicts can be adjusted according to actual business needs).

[0074] Channel sharing: Using the same channel and transmitting data at the same time in the same time window.

[0075] Step 4: Edge Weight Calculation and Graph Construction

[0076] Edge weight calculation: w ij =α1*interference intensity +α2*conflict frequency +α3*time correlation

[0077] Interference intensity: Based on RSSI difference and distance attenuation model, normalized to [0,1].

[0078] Conflict frequency: Number of conflicts per unit time, reflected after normalization.

[0079] Time correlation: Time synchronization of state changes of two devices (such as the frequency of simultaneous channel switching).

[0080] Parameter weights α1, α1 and α3 can be adjusted through experiments (such as default values 0.4, 0.3, 0.3).

[0081] Construct undirected weighted conflict graph: Combine nodes and edges to generate an initial graph structure.

[0082] Step 5: Dynamic Graph Structure Update

[0083] Trigger scenarios: Device access / exit, channel switching, power adjustment, and other state changes.

[0084] Update mode: add / delete nodes and edges, or recalculate existing edge weights (e.g., update related edge weights when devices switch channels).

[0085] The spatio-temporal conflict graph constructed from the specific embodiment 2 needs to be further analyzed for the conflict possibility between devices. The conflict modeling process based on the graph attention network (GAT) is shown in the specific embodiment 3. The conflict modeling process based on the graph attention network (GAT) in the specific embodiment 3 is as follows.

[0086] Step 1: input conflict graph node feature matrix

[0087] The node feature vector generated in embodiment 2 is input, including device normalized coordinates, channel encoding, and conflict times in the time window, etc. spatio-temporal fusion information in the form of a matrix (node number x feature number) is input into the model.

[0088] Step 2: multi-layer GAT layer calculation and feature extraction

[0089] The conflict graph is processed through a multi-layer graph attention network:

[0090] Linear transformation: use a weight matrix to increase the dimension of node features (e.g., 10D→32D), enhance the feature expression ability.

[0091] Attention calculation: calculate the correlation of nodes and their neighbors, generate attention weights through softmax normalization, and represent the importance of node interaction.

[0092] Feature aggregation: weighted aggregation of neighbor node features and activation, capturing complex relationships between devices layer by layer, outputting low-dimensional features (e.g., 16D) containing high-order dependencies.

[0093] Step 3: fully connected layer mapping and conflict probability prediction

[0094] Map the low-dimensional features to 1D through the fully connected layer, output the conflict probability P through the Sigmoid activation, reflect the conflict possibility of the device in the next time slot, and provide decision basis for the optimization of the backoff strategy.

[0095] The conflict probability output from the specific embodiment 3 needs to be further used for backoff strategy optimization. The backoff strategy optimization process based on multi-agent reinforcement learning is shown in the specific embodiment 4. The multi-agent reinforcement learning (MADDPG) backoff strategy optimization process in the specific embodiment 4 is as follows.

[0096] Step 1: agent state definition and initialization

[0097] Each wireless device is regarded as an independent agent, and the state space contains the device's own characteristics (such as normalized coordinates, current channel number, and conflict probability prediction value) and neighbor node association information. The initialized agent state and policy network, value network parameters are initialized.

[0098] Step 2: Action execution and environment interaction

[0099] The agent selects an action (such as adjusting the backoff window size, switching channels, or changing the transmission power) according to the current policy network. After executing the action, the agent obtains environmental feedback, including the reward value (such as successful transmission + 10, conflict - 5) and the new state.

[0100] Step 3: Experience collection and training data generation

[0101] The agent's state, action, reward, and next state are stored in the experience buffer. When the buffer data reaches a threshold (such as 1000), a batch of data is randomly selected for training.

[0102] Step 4: Network training and policy update

[0103] The network parameters are updated through the multi-agent deep deterministic policy gradient (MADDPG) algorithm:

[0104] Value network (Critic): Input state and action, evaluate action value, optimization goal is to minimize value estimation error.

[0105] Policy network (Actor): Output optimal action according to state, maximize expected reward through gradient ascent.

[0106] Target network soft update: Regularly synchronize the main network parameters to the target network slowly, stabilize the training process.

[0107] Step 5: Optimal policy execution

[0108] When the number of training rounds reaches a preset value (such as 10000 rounds) or the performance indicator meets the standard, the agent executes the backoff action according to the optimal policy, achieving conflict probability reduction and spectrum utilization improvement.

[0109] The optimal backoff strategy output from the specific embodiment 4 needs to be continuously optimized in combination with the real-time network state. The system optimization process based on dynamic feedback is shown in specific embodiment 5. Specific embodiment 5: Dynamic feedback and system optimization process

[0110] Step 1: Real-time performance monitoring indicators

[0111] Monitoring content: Through the monitoring module deployed in the data processing center, the following indicators are collected in real time:

[0112] Conflict rate: Statistics of the number of conflicts per 100 time slots, reflecting the frequency of device access conflicts.

[0113] Spectrum utilization: Calculate the proportion of the amount of data successfully transmitted in a unit of time to the total available channel resources.

[0114] Throughput: Measure the amount of effective data transmitted by the system in a unit of time.

[0115] Device status: Monitor device access / removal events, current working channel distribution, transmit power changes, etc.

[0116] Data source: Wireless devices periodically report their status (such as sending a status package every 100ms) through the built-in communication module, combined with historical interaction data in the conflict graph for comprehensive analysis.

[0117] Step 2: Trigger condition judgment

[0118] Determine whether the conditions for triggering the decision are met. Any one of the policy adjustment or model update conditions is met. If met, proceed to step 3; otherwise, return to step 1.

[0119] Policy adjustment trigger conditions (meet any one of the following):

[0120] Conflict rate exceeds the preset threshold for 5 consecutive minutes (such as 15 times per 100 time slots);

[0121] Spectrum utilization is less than 80% of the target value (such as the target value is 50%, the trigger threshold is 40%);

[0122] Throughput is less than the set minimum value for 10 consecutive minutes (such as 80Mbps).

[0123] Model update trigger conditions (meet any one of the following):

[0124] Device quantity variation amplitude ≥10% (such as originally 50 devices, add / remove ≥5 devices);

[0125] Significant changes in channel interference patterns are detected (such as continuous high-intensity interference signals in the same channel).

[0126] Step 3: Lightweight policy condition judgment.

[0127] Applicable scenarios for lightweight policy challenges: Small changes in devices (such as adding ≤5 devices) or short-term performance fluctuations, without the need to reconstruct the conflict graph.

[0128] Determine whether the lightweight policy adjustment conditions are met, and if so, proceed to step 4; otherwise, proceed to step 5.

[0129] Step 4: Lightweight policy adjustment

[0130] Execution process:

[0131] Clear the experience buffer: Delete historical training data to prevent outdated information from affecting new strategy learning;

[0132] Restart MADDPG training: Reinitialize the multi-agent reinforcement learning process based on the existing conflict graph structure;

[0133] Parameter optimization: Only the policy network (Actor) and value network (Critic) parameters are updated, preserving the GAT model's conflict graph feature extraction capabilities;

[0134] Incremental training: Set a limited number of training rounds (such as 5000 rounds) to quickly converge to a suboptimal strategy that adapts to the new environment.

[0135] Step 5: Full model update

[0136] Applicable scenarios: large-scale device changes (such as adding ≥10 devices) or channel topology reconstruction (such as changes in frequency band occupancy patterns).

[0137] Execution process:

[0138] Conflict graph reconstruction:

[0139] Re-collect multi-dimensional information of all devices (coordinates, channels, conflict history, etc.);

[0140] Generate node features based on new data, regenerate edges according to the rule of "distance < threshold, historical conflicts ≥ N times, and co-channel transmission", and calculate edge weights (formula: w ij =α1*interference intensity+α2*conflict frequency+α3*time correlation).

[0141] GAT model retraining:

[0142] Input the node feature matrix of the new conflict graph, retrain the multi-layer graph attention network, and update the attention weights and fully connected layer parameters;

[0143] The goal is to relearn the spatiotemporal correlation features between devices and output accurate collision probability prediction values.

[0144] MADDPG full training:

[0145] Based on the new conflict graph and the features of GAT output, the agent state space is redefined;

[0146] Execute the complete multi-agent reinforcement learning process (initialize the network → collect experience → train and update → policy convergence) and learn a backoff strategy that adapts to the new network topology.

[0147] Step 6: Optimization effect verification

[0148] Verification indicators:

[0149] Whether the conflict rate is reduced below the threshold value (such as ≤10 times per 100 time slots);

[0150] Whether the spectrum utilization is improved to the target range (such as ≥55%);

[0151] Whether the throughput is restored to the set value (such as ≥90 Mbps).

[0152] Iteration mechanism:

[0153] If the indicators after optimization do not meet the standards, automatically restart step 3 or step 4 (select light or full update according to trigger conditions);

[0154] If the standards are met, enter the normal monitoring mode, and regularly (such as every hour) perform light parameter fine-tuning to maintain system performance stability.

[0155] The block diagram of the unlicensed band conflict modeling and backoff optimization device based on graph neural network is shown in Figure 6 It includes data acquisition unit, data processing unit, graph construction unit, conflict modeling unit, strategy optimization unit, and feedback optimization unit. The functions of each unit module are as follows:

[0156] Data acquisition unit, for acquiring multi-dimensional information of wireless devices;

[0157] Data processing unit, for preprocessing the collected data and extracting spatio-temporal features;

[0158] Graph construction unit, for constructing a conflict graph associated with time and space;

[0159] Conflict modeling unit, for processing the conflict graph based on graph attention network;

[0160] Strategy optimization unit, for optimizing backoff strategies using multi-agent reinforcement learning;

[0161] Feedback optimization unit, for real-time monitoring of system performance and updating strategy and model parameters.

[0162] The relationship between each unit module is explained as follows:

[0163] Data acquisition unit: responsible for acquiring multi-dimensional information of wireless devices (such as device identification, coordinates, channel number, etc.), and transmitting raw data to the data processing unit.

[0164] Data processing unit: preprocess the raw data (normalization, encoding, spatio-temporal feature extraction), generate feature vectors containing spatio-temporal correlation, and output to graph construction unit.

[0165] Graph construction unit: abstract the device as a graph node, generate edges and weights according to the distance between devices, historical conflicts, channel sharing, etc., construct a conflict graph, and output to the conflict modeling unit.

[0166] Conflict modeling unit: analyze the conflict graph based on graph attention network (GAT), output the device conflict probability, and use it to generate the backoff strategy for the strategy optimization unit.

[0167] Strategy optimization unit: generate the optimal backoff strategy (such as adjusting the backoff window, switching channels) using multi-agent deep deterministic policy gradient (MADDPG), and feed back the strategy execution result to the feedback optimization unit.

[0168] Feedback optimization unit: real-time monitor system performance (such as conflict rate, spectrum utilization), trigger parameter update (such as re-preprocess data, re-construct conflict graph, re-train model, etc.) according to the index, and form a closed-loop optimization.

[0169] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited to this. Any skilled person in the art can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present application, and these modifications or replacements should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for modeling and backoff optimization of unlicensed frequency band conflicts based on graph neural networks, characterized in that: The following steps are involved: The data acquisition module obtains multi-dimensional information of wireless devices, including the device's unique identification, geographic location coordinates, transmission power, transmission rate, current working channel number, communication status of each time slot, and timestamp; The data processing module pre-processes the collected data, including removing abnormal data, normalizing continuous data, encoding discrete data, dividing data by time windows, extracting time and space dimension features, and fusing them to form device feature vectors containing time and space correlations; The construction module defines each wireless device as a graph node, generates edges based on the distance between devices, historical conflicts, and channel sharing, calculates edge weights, and constructs a spatiotemporal conflict graph. The conflict graph is processed based on the graph attention network, and node features containing complex device associations are output through multi-layer calculations to predict the probability of device conflicts. Each wireless device is regarded as an intelligent agent, and a multi-agent deep deterministic policy gradient algorithm is used to learn the backoff strategy. Each intelligent agent performs the optimal backoff action based on the current state. Monitor device communication status and system performance indicators in real time. If the performance does not meet expectations, re-trigger the multi-agent reinforcement learning to update the backoff strategy. Update the conflict graph structure and graph neural network model parameters regularly or when the network topology changes significantly.

2. The method according to claim 1, characterized in that The data acquisition module periodically collects basic attribute information of the device through the built-in communication module of the wireless device, and records the communication status information of the device in real time.

3. The method according to claim 1, characterized in that The data processing module uses a maximum-minimum normalization method to map continuous data to the [0,1] interval.

4. The method according to claim 1, characterized in that If the distance between two devices is less than the preset interference threshold, or multiple conflicts have occurred in multiple consecutive time windows, or they share the same working channel and have data transmission behavior in the same time window, an edge is added between the corresponding nodes.

5. The method according to claim 4, characterized in that The calculation of the edge weights also considers the degree of time synchronization of device state changes, which is reflected by the time correlation parameter.

6. The method according to claim 1, characterized in that The graph attention network sets up multiple layers of attention mechanism layers and multiple attention heads, and adaptively learns the importance between nodes through attention calculation.

7. The method according to claim 6, characterized in that The conflict graph is processed based on the graph attention network, including: linear transformation of node features, calculation and normalization of attention coefficients between nodes, and weighted aggregation of neighbor node features to update the current node features.

8. The method according to claim 1, characterized in that In the multi-agent deep deterministic policy gradient algorithm, each agent maintains a policy network and a value network, and updates the network parameters by sampling the state-action sequence of the agent.

9. A device for modeling and backoff optimization of unlicensed frequency band conflicts based on graph neural networks, characterized in that: include: A data acquisition unit, used to obtain multi-dimensional information of wireless devices; A data processing unit, used to pre-process the collected data and extract spatiotemporal features; A graph construction unit, used to construct a conflict graph with temporal and spatial associations; Conflict modeling unit, used to process the conflict graph based on the graph attention network; A policy optimization unit, used to optimize the backoff strategy using multi-agent reinforcement learning; Feedback optimization unit, used to monitor system performance in real time and update strategy and model parameters.

10. A computer storage medium, characterized in that The storage medium stores program code, and when the program code is executed by a computing device, the method according to any one of claims 1 to 8 is executed.

Images