A method for airborne WiFi network resource allocation based on spatiotemporal graph neural networks

By constructing a heterogeneous graph of the airborne cabin WiFi network and using a spatiotemporal graph neural network model, the network resource demand and risk status are predicted, and resource allocation strategies are generated and optimized. This solves the problems of congestion and low resource utilization that cannot be predicted in advance in traditional methods, and realizes proactive prediction and joint scheduling of multiple devices in the airborne WiFi network, thereby improving system stability and resource utilization.

CN123123945APending Publication Date: 2026-11-17SHANGHAI COKER AVIATION TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611268771.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-20
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Traditional airborne WiFi network resource configuration methods cannot predict network congestion trends in advance, cannot achieve lightweight intelligent decision-making in airborne edge computing environments, and cannot effectively characterize the coupling relationship between multi-level devices and links, resulting in passive response and low resource utilization when the network is congested.

Method used

A spatiotemporal graph neural network-based approach is used to construct a heterogeneous graph of the airborne cabin WiFi network. The spatiotemporal graph neural network model is used to predict network resource requirements and risk status, generate candidate strategies for control actions of various devices, and conduct a comprehensive evaluation to select the optimal configuration strategy.

Benefits of technology

It enables proactive prediction and multi-device joint scheduling of airborne WiFi networks, reducing network lag and improving resource utilization and system stability.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123123945A_ABST
    Figure CN123123945A_ABST
Patent Text Reader

Abstract

This invention relates to the field of aviation communication network resource management technology, and discloses a method for configuring airborne WiFi network resources based on a spatiotemporal graph neural network. The method includes: constructing a heterogeneous graph of the airborne cabin WiFi network, representing the network state as a heterogeneous graph containing multiple types of device nodes and various connection edges; collecting network states within a continuous time window to form a dynamic graph sequence; inputting the sequence into a spatiotemporal graph neural network model to predict link congestion risk, AP load, quality of service, and content requirements; generating a candidate strategy set containing multiple device control actions based on the prediction results; comprehensively evaluating each candidate strategy; selecting the optimal strategy that meets the constraints; issuing and executing the optimal strategy; and providing feedback for correction. This invention achieves dynamic configuration of airborne WiFi network resources and is suitable for local intelligent scheduling in airborne environments.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of aviation communication network resource management technology, specifically to an airborne WiFi network resource configuration method based on a spatiotemporal graph neural network. Background Technology

[0002] With the development of in-flight wireless LANs and in-flight entertainment systems, passengers can access cabin WiFi networks via personal mobile terminals or seat display terminals to access content such as video-on-demand, music, e-reading, flight information, shopping services, and destination services. Compared to wireless networks in ground offices, hotels, or shopping malls, in-flight cabin WiFi networks are characterized by enclosed spaces, user distribution constrained by seating, overlapping AP coverage areas, and limited link bandwidth and local computing power. When a large number of passengers simultaneously access multimedia services, problems such as excessive load on local APs, backbone link congestion, media service unit request queuing, and terminal content synchronization conflicts can easily occur simultaneously, leading to increased video first-frame waiting time, playback stuttering, slow webpage response, and impact on system management services.

[0003] Traditional resource allocation methods typically employ fixed parameters or threshold triggering, such as pre-setting the maximum number of AP connections, port rate limiting ratios, and service queue priorities. Adjustments are only implemented when link utilization or the number of AP connections exceeds the threshold. While simple to implement, these methods have significant shortcomings: First, they cannot predict network congestion trends in advance, leading to reactive measures only after problems have occurred. Second, they focus on individual devices or links, failing to characterize the multi-level coupling relationships between MSUs, switches, ACs, APs, SDUs, and passenger terminals. Third, they lack joint optimization between wireless access control, switching link scheduling, and media content distribution, with each control component operating independently. Fourth, during flight, continuous real-time scheduling via the ground cloud is often unreliable, requiring local intelligent inference on the airborne side, while existing methods lack lightweight intelligent decision-making mechanisms suitable for airborne edge computing environments. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an airborne WiFi network resource allocation method based on a spatiotemporal graph neural network. This method has the advantages of proactively predicting airborne WiFi network congestion and jointly scheduling multiple devices through the spatiotemporal graph neural network, effectively reducing lag, improving resource utilization and system stability, and solving the aforementioned technical problems.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for configuring airborne WiFi network resources based on a spatiotemporal graph neural network, comprising the following steps: S1: Construct a heterogeneous graph of the airborne cabin WiFi network, representing the network state as a heterogeneous graph containing multiple types of device nodes and various connection edges; S2: Collect network states within a continuous time window to form a dynamic graph sequence, input it into a spatiotemporal graph neural network model, and predict network resource requirements and risk status at future moments; S3: Based on the prediction results of S2, generate a set of candidate resource configuration strategies that include multiple types of device control actions; S4: Perform a comprehensive evaluation on each candidate strategy generated in S3, and select the strategy that satisfies the constraints and has the smallest evaluation function value as the target resource allocation strategy. S5: Send the target resource configuration strategy selected in S4 to the corresponding device for execution, and make feedback and corrections based on the execution results.

[0006] As a preferred technical solution of the present invention, the S1 heterogeneous graph includes a set of nodes, a set of edges, node features, and edge features; The expression for the heterogeneous graph is as follows: ; in, Represents a set of device nodes. Represents the set of connecting edges. Represents the node state feature matrix. Represents the edge state feature matrix; The node set consists of multiple types of device nodes and user service nodes, and the expression of the node set is as follows: ; in, This indicates the airborne media service unit node. Indicates a switch node. Indicates a wireless controller node. Indicates a wireless access point node. Indicates the terminal unit node. This indicates a passenger mobile terminal node or a user aggregation node; The expression for the set of connecting edges is as follows: ; in, This indicates the backbone link between the MSU and the switch. This indicates the access link between the switch and the AC, AP, or SDU. This indicates the wireless connection between the AP and the passenger terminal. This indicates the logical business relationship for distributing media content from the MSU to users or terminals. This indicates a cascaded link or backup link between SDUs; The expression for the node feature is as follows: ; in, This means that for any node At any given moment The node feature vectors, For node type, For processor utilization, For memory usage, For business traffic, For the length of the request queue, For the number of connected users, In cached state, The device is either online or in an abnormal state; The expression for the edge feature is as follows: ; in, Represents any connecting edge The edge feature vectors, Indicates the link type. Indicates the nominal bandwidth. Indicates real-time utilization rate. Indicates transmission delay. Indicates packet loss rate. Indicates latency jitter. This indicates the error or anomaly count.

[0007] As a preferred embodiment of the present invention, the expression of the S2 dynamic graph sequence is as follows: ; in, Indicates the current scheduling time A dynamic graph sequence consisting of network state graphs at L consecutive sampling times, with the endpoint being the endpoint. , and These represent the heterogeneous network state diagrams corresponding to the start time t-L+1, the next sampling time t-L+2, and the current time t, respectively, where L represents the number of historical sampling steps included in the input time window. The S2 prediction of future network resource demands and risk states is based on the spatiotemporal graph neural network model given... Predicting the future under certain conditions The network resource demand and risk status within each time step are expressed as follows: ; in, This represents the joint multi-task prediction results from time t+1 to t+H, including link congestion probability, AP load, QoS risk, and content requirements. H represents the number of future time steps included in the prediction time span. This represents a spatiotemporal graph neural network mapping function that takes a dynamic graph sequence as input and outputs the future network resource demand and risk status. The spatiotemporal graph neural network model consists of a heterogeneous feature encoding layer, a spatial graph attention layer, a time series feature extraction layer, a multi-task prediction output layer, and a confidence evaluation layer. The relevant expression for the heterogeneous feature encoding layer is as follows: ; in, Represents a node At any moment The initial hidden features, This represents the encoding network corresponding to the node type. This means that for any node At any given moment Node feature vectors; The relevant expression for the spatial graph attention layer is as follows: ; in, Represents a node No. Spatial features after layer update It is a non-linear activation function. and For a trainable parameter matrix, Represents a node No. Spatial characteristics of layers Represents a node No. Spatial characteristics of layers For nodes For nodes Attention weights Represents a set of edge types. Indicates passing through the first Edges and Nodes The set of connected neighbor nodes, and This represents the summation operation; The relevant expressions for the time series feature extraction layer are as follows: ; in, Represents a node The spatiotemporal fusion representation at the current scheduling time t, This represents a gated recurrent unit network used to extract the temporal dependencies of node states. , and These represent the spatial features of node v at the beginning time t-L+1, subsequent time t-L+2, and current time t of the input time window, respectively, after passing through the spatial graph attention layer, where L represents the length of the time window. The multi-task prediction output layer is used to simultaneously predict link congestion, AP load, QoS risk, and content demand. The prediction expression for link congestion is as follows: ; in, Indicates connecting edges The probability of congestion occurring within a future predicted time period. and Let i and j represent the spatiotemporal fusion representations of the two endpoints of the edge at the current time t, respectively. This represents the edge feature vector connecting edge (i,j) at the current time t. This represents a link prediction decoder that maps node representations and edge features to link congestion probabilities. The prediction expression for the AP load is as follows: ; in, This indicates the target AP's load rate or the expected number of connected users over a future forecast period. This represents the spatiotemporal fusion representation of the target AP at the current time t. This refers to an AP load prediction decoder that maps the spatiotemporal fusion representation of APs to future load rates or the number of connected users. The relevant expression for the QOS risk is as follows: ; in, This indicates the probability of media playback stuttering, increased first-frame waiting time, or decreased service quality. This represents the spatiotemporal fusion representation of the target media service or service aggregation node at the current time t. This represents a QoS risk prediction decoder that maps the spatiotemporal fusion representation of business to the probability of service quality degradation. The relevant expressions for the content requirements are as follows: ; in, This indicates the expected access demand for the specified content within a future time period. This represents the spatiotemporal fusion representation of the airborne media service unit node at the current time t. This represents the spatiotemporal fusion representation of the target media service or user service aggregation node related to the specified content at the current time t. This refers to a content demand prediction decoder that predicts future access needs for specified content based on MSU status and business demand characteristics. The confidence evaluation layer is used to improve the accuracy of link load prediction, AP load prediction, QoS risk classification, and content demand prediction. The relevant expressions are as follows: ; in, This represents the value of the comprehensive loss function; To account for errors in link utilization and congestion risk prediction, Errors in predicting the number of AP users or load rate. The classification error is due to a decline in the quality of media services. For content demand prediction error, To ensure compliance with penalties for violations of business resource constraints, , , , and These are the weighting coefficients for each loss term.

[0008] As a preferred embodiment of the present invention, the expression for the S3 candidate resource configuration strategy set is as follows: ; in, This represents the set of candidate resource allocation strategies generated at time t. , and These represent the 1st, 2nd, and nth candidate resource configuration policies in the set, respectively, where n represents the total number of candidate policies. Each candidate policy consists of one or more control actions from the switch, AP / AC, MSU, and SDU.

[0009] As a preferred embodiment of the present invention, the expression for the S4 comprehensive evaluation is as follows: ; in, This represents the evaluation function value. Indicate candidate strategies, This represents the predicted link congestion cost after implementing the strategy. This indicates the predicted cost of media buffering. This indicates the cost of uneven load distribution among the access points (APs). This indicates the stability cost corresponding to the magnitude and frequency of configuration adjustments. This indicates the risk and cost of impacting service security or causing link unavailability. , , , and Weights for evaluation indicators; The expression for the S4 constraint is as follows: ; in, This indicates the actual bandwidth obtained by the guaranteed service. This indicates the minimum bandwidth threshold required to guarantee service availability. The expression for selecting the strategy that satisfies the constraints and minimizes the evaluation function value as the target resource allocation strategy is as follows: ; in, This represents the objective resource allocation strategy that satisfies all constraints and minimizes the evaluation function value. This represents the set of candidate resource allocation strategies at time t. This represents the minimization selection operation, which searches for and returns the strategy that minimizes the evaluation function value from the set of candidate resource allocation strategies.

[0010] Compared with existing technologies, this invention provides an airborne WiFi network resource allocation method based on spatiotemporal graph neural networks, which has the following beneficial effects: 1. This invention models the airborne cabin WiFi network as a heterogeneous dynamic graph containing multiple types of device nodes such as MSU, switches, AC, AP, SDU, and passenger terminals, as well as various connection edges such as backbone links, access links, wireless links, service distribution relationships, and SDU cascade links. This fully depicts the physical connection relationships and logical service transmission relationships between multiple devices and links in the airborne network, overcoming the limitations of existing technologies that only focus on a single device or a single link, and laying the model foundation for subsequent accurate prediction and joint scheduling.

[0011] 2. This invention maps the original features of different types of devices to a unified vector space through the heterogeneous feature encoding layer in the spatiotemporal graph neural network model. The spatial graph attention layer performs weighted aggregation of neighbor node information based on the heterogeneous graph connection relationship to capture the degree of spatial influence between different devices and links. The time series feature extraction layer extracts the time evolution trend of load and service requests. The multi-task prediction output layer simultaneously outputs four prediction results: link congestion risk, AP load, service quality risk, and content demand. It can identify risks in advance before congestion occurs, overcoming the technical defect of existing threshold triggering methods that can only respond passively after the problem occurs. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of the overall structure of the airborne cabin WiFi network of the present invention; Figure 2 This is a schematic diagram of the process of the present invention. Detailed Implementation

[0013] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0014] Please see Figures 1-2 A method for configuring airborne WiFi network resources based on spatiotemporal graph neural networks includes the following steps: S1: Construct a heterogeneous graph of the airborne cabin WiFi network, representing the network state as a heterogeneous graph containing multiple types of device nodes and various connection edges; S1 heterogeneous graph includes a set of nodes, a set of edges, node features, and edge features; The expression for a heterogeneous graph is as follows: ; in, Represents a set of device nodes. Represents the set of connecting edges. Represents the node state feature matrix. The edge state feature matrix is ​​represented. Since airborne networks contain multiple types of devices and links, this graph model is a heterogeneous dynamic graph, which can express both physical connection relationships and logical service transmission relationships. The node set consists of multiple types of device nodes and user service nodes. The expression for the node set is as follows: ; in, This indicates the airborne media service unit node. Indicates a switch node. Indicates a wireless controller node. Indicates a wireless access point node. Indicates the terminal unit node. This indicates a passenger mobile terminal node or a user aggregation node; In actual deployment, if each passenger terminal is modeled separately, the graph size will grow rapidly with the number of users. To reduce the pressure of onboard local inference, terminals can be aggregated into user service nodes according to AP, cabin segment, service type, or seat area, such as "front cabin video user aggregation node", "middle cabin ordinary browsing user aggregation node", or "AP1 attached video user node". This aggregation method not only preserves user distribution and service type information, but also reduces the computational complexity of the model. The expression for the set of connecting edges is as follows: ; in, This indicates the backbone link between the MSU and the switch. This indicates the access link between the switch and the AC, AP, or SDU. This indicates the wireless connection between the AP and the passenger terminal. This indicates the logical business relationship for distributing media content from the MSU to users or terminals. This indicates a cascaded link or backup link between SDUs; By classifying the edges as described above, the model can distinguish between physical links, wireless access relationships, media service relationships, and terminal cascading relationships, thereby identifying the impact of different types of connections on congestion propagation and changes in service quality. The expression for node features is as follows: ; in, This means that for any node At any given moment The node feature vectors, For node type, For processor utilization, For memory usage, For business traffic, For the length of the request queue, For the number of connected users, In cached state, This indicates whether the device is online or in an abnormal state; different types of nodes can have different original parameters, which are subsequently mapped to a hidden feature space of the same dimension by the corresponding coding layer of the type; The expression for the edge feature is as follows: ; in, Represents any connecting edge The edge feature vectors, Indicates the link type. Indicates the nominal bandwidth. Indicates real-time utilization rate. Indicates transmission delay. Indicates packet loss rate. Indicates latency jitter. The network status data, representing bit error or anomaly counts, is shown in Table 1. Table 1. In-flight cabin WiFi network status data

[0015] S2: Collect network states within a continuous time window to form a dynamic graph sequence, input it into a spatiotemporal graph neural network model, and predict network resource requirements and risk status at future moments; The expression for the S2 dynamic graph sequence is as follows: ; in, Indicates the current scheduling time A dynamic graph sequence consisting of network state graphs at L consecutive sampling times, with the endpoint being the endpoint. , and These represent the heterogeneous network state diagrams corresponding to the start time t-L+1, the next sampling time t-L+2, and the current time t, respectively, where L represents the number of historical sampling steps included in the input time window. S2 predicts future network resource demands and risk states using a spatiotemporal graph neural network model given... Predicting the future under certain conditions The network resource demand and risk status within each time step are expressed as follows: ; in, This represents the joint multi-task prediction results from time t+1 to t+H, including link congestion probability, AP load, QoS risk, and content requirements. H represents the number of future time steps included in the prediction time span. This represents a spatiotemporal graph neural network mapping function that takes a dynamic graph sequence as input and outputs the future network resource demand and risk status. The overall model consists of a heterogeneous feature encoding layer, a spatial graph attention layer, a time series feature extraction layer, a multi-task prediction output layer, and a confidence evaluation layer. The heterogeneous feature encoding layer is responsible for mapping the raw features of different types of devices to a unified vector space; the spatial graph attention layer is responsible for modeling the resource impact between adjacent nodes; the time series feature extraction layer is responsible for capturing the changing trends of load and service requests; and the multi-task output layer predicts link congestion, AP load, QoS risk, and content requirements, respectively. The spatiotemporal graph neural network model consists of a heterogeneous feature encoding layer, a spatial graph attention layer, a time series feature extraction layer, a multi-task prediction output layer, and a confidence evaluation layer. The heterogeneous feature encoding layer uses a type-related encoder to perform feature mapping on different nodes, and the relevant expression is as follows: ; in, Represents a node At any moment The initial hidden features, This represents the encoding network corresponding to the node type. This means that for any node At any given moment The encoder focuses on the node feature vectors; for MSU nodes, the encoder focuses on CPU, storage read rate, media concurrent stream and cache hit rate; for AP nodes, the encoder focuses on the number of access users, channel utilization, RSSI, SNR and retransmission rate; for switch nodes, the encoder focuses on port utilization, queue length and packet loss rate; for SDU nodes, the encoder focuses on synchronization rate, heartbeat status and cascade link status. The spatial graph attention layer performs weighted aggregation of neighbor node information based on the connection relationships in the heterogeneous graph. The relevant expression is as follows: ; in, Represents a node No. Spatial features after layer update It is a non-linear activation function. and For a trainable parameter matrix, Represents a node No. Spatial characteristics of layers Represents a node No. Spatial characteristics of layers For nodes For nodes Attention weights Represents a set of edge types. Indicates passing through the first Edges and Nodes The set of connected neighbor nodes, and This represents the summation operation; Attention weights are used to measure the degree of influence of different neighboring nodes on the current state of the target node. For example, when the number of concurrent media streams of an MSU continues to rise, the attention weights of the backbone links and switch nodes connected to it may increase. When the number of users and retransmission rate of an AP increases, the impact of neighboring APs and upstream switching links on its load state will be further considered by the model. After obtaining the spatial graph features at each time step, the model further employs time series structures such as GRU, LSTM, TCN, or Transformer to extract the load change trend; taking GRU as an example, the relevant expression for the time series feature extraction layer is as follows: ; in, Represents a node The spatiotemporal fusion representation at the current scheduling time t, This represents a gated recurrent unit network used to extract the temporal dependencies of node states. , and These represent the spatial features of node v at the start time t-L+1, subsequent times t-L+2, and the current time t, respectively, output by the spatial graph attention layer, where L represents the length of the time window. This representation includes not only the current network topology information but also the evolution trend of user access, link utilization, media request volume, and anomaly indicators over a past period. The multi-task prediction output layer is used to simultaneously predict link congestion, AP load, QoS risk, and content demand. The prediction expression for link congestion is as follows: ; in, Indicates connecting edges The probability of congestion occurring within a future predicted time period. and Let i and j represent the spatiotemporal fusion representations of the two endpoints of the edge at the current time t, respectively. This represents the edge feature vector connecting edge (i,j) at the current time t. This represents a link prediction decoder that maps node representations and edge features to link congestion probabilities. The prediction expression for AP load is as follows: ; in, This indicates the target AP's load rate or the expected number of connected users over a future forecast period. This represents the spatiotemporal fusion representation of the target AP at the current time t. This refers to an AP load prediction decoder that maps the spatiotemporal fusion representation of APs to future load rates or the number of connected users. The relevant expression for QOS risk is as follows: ; in, This indicates the probability of media playback stuttering, increased first-frame waiting time, or decreased service quality. This represents the spatiotemporal fusion representation of the target media service or service aggregation node at the current time t. This represents a QoS risk prediction decoder that maps the spatiotemporal fusion representation of business to the probability of service quality degradation. The relevant expressions for content requirements are as follows: ; in, This indicates the expected access demand for the specified content within a future time period. This represents the spatiotemporal fusion representation of the airborne media service unit node at the current time t. This represents the spatiotemporal fusion representation of the target media service or user service aggregation node related to the specified content at the current time t. This refers to a content demand prediction decoder that predicts future access needs for specified content based on MSU status and business demand characteristics. The confidence evaluation layer is used to improve the accuracy of link load prediction, AP load prediction, QoS risk classification, and content demand prediction. The relevant expressions are as follows: ; in, This represents the value of the comprehensive loss function; To account for errors in link utilization and congestion risk prediction, Errors in predicting the number of AP users or load rate. The classification error is due to a decline in the quality of media services. For content demand prediction error, To ensure compliance with penalties for violations of business resource constraints, , , , and These are the weighting coefficients for each loss term; S3: Based on the prediction results of S2, generate a set of candidate resource configuration strategies that include multiple types of device control actions, and its expression is as follows: ; in, This represents the set of candidate resource allocation strategies generated at time t. , and These represent the 1st, 2nd, and nth candidate resource configuration policies in the set, respectively, where n represents the total number of candidate policies. Each candidate policy consists of one or more braking actions from the switch, AP / AC, MSU, and SDU. S4: A comprehensive evaluation is performed on each candidate strategy generated in S3, and the strategy that satisfies the constraints and has the smallest evaluation function value is selected as the target resource configuration strategy. Candidate strategies consist of a combination of various device control actions, including switch queue priority adjustment, port bandwidth quota adjustment, AP user access limit adjustment, new user access guidance, AP channel or power adjustment, MSU popular content pre-caching, media bitrate adjustment, SDU packet synchronization, and backup path switching, as shown in Table 2. Table 2 Resource Allocation Strategies and Control Objects

[0016] For each candidate strategy The system performs a comprehensive evaluation from the perspectives of congestion risk, media stuttering, load balancing, configuration stability, and security risks. The expression for the comprehensive evaluation is as follows: ; in, This represents the evaluation function value. Indicate candidate strategies, This represents the predicted link congestion cost after implementing the strategy. This indicates the predicted cost of media buffering. This indicates the cost of uneven load distribution among the access points (APs). This indicates the stability cost corresponding to the magnitude and frequency of configuration adjustments. This indicates the risk and cost of impacting service security or causing link unavailability. , , , and Weights for evaluation indicators; The safeguard business constraint is used to ensure that system management communications, equipment status monitoring communications, cabin service announcements related data, and network operation and maintenance data are not affected by ordinary entertainment services; its minimum bandwidth requirement can be expressed as: ; in, This indicates the actual bandwidth obtained by the guaranteed service. This indicates the minimum bandwidth threshold required to guarantee service availability. Stability constraints are configured to prevent the model from frequently adjusting network parameters due to short-term fluctuations; the system can set the minimum interval for AP access guidance, the minimum interval for channel switching, the maximum single adjustment range of bandwidth configuration, the maximum single change ratio of media bitrate, and the minimum observation time after the policy takes effect. Safety rollback constraints are used to handle situations where the model is unavailable or the output is unreliable. When the model inference module is abnormal, the missing input data rate exceeds the threshold, the model output confidence is lower than the threshold, the network performance deteriorates significantly after the policy is executed, or multiple key devices are abnormal at the same time, the system switches to static rule or threshold rule control mode to ensure the availability of basic services. The expression for selecting the strategy that satisfies the constraints and minimizes the evaluation function value as the target resource allocation strategy is as follows: ; in, This represents the objective resource allocation strategy that satisfies all constraints and minimizes the evaluation function value. This represents the set of candidate resource allocation strategies at time t. This represents the minimization selection operation, which searches for and returns the strategy that minimizes the evaluation function value in the set of candidate resource allocation strategies. S5: Send the target resource configuration strategy selected in S4 to the corresponding device for execution, and make feedback and corrections based on the execution results; The system implementation process is as follows: Figure 2 As shown, the process involves the following steps: First, the data acquisition module collects network status data from the MSU, SW, AC, AP, SDU, and user terminal sides. Second, the topology construction module generates a heterogeneous dynamic graph based on physical connections and logical service relationships. Third, the prediction module inputs the graph sequence within a continuous time window into a spatiotemporal graph neural network and outputs risk prediction results. Subsequently, the policy generation module generates candidate configuration policies, and the constraint decision module filters the target policies. Finally, the execution module distributes the policies to the corresponding devices, and the feedback optimization module adjusts the model or policy parameters based on the execution effect. To reduce system risks during flight, this invention recommends a deployment approach of "local inference, lightweight correction, and post-flight training." That is, during flight, model inference and policy execution are completed by the MSU or airborne edge computing devices without large-scale online training. For prediction bias, only the congestion threshold, policy weight, or confidence threshold is corrected. After the flight, the de-identified or aggregated operational data is synchronized to the ground platform for model retraining and version updates. During the cruise phase, a large number of passengers simultaneously connect to the cabin WiFi and access popular movies. The system detects a continuous increase in the number of concurrent media streams from the MSU, an increase in the utilization rate of the backbone link from the MSU to the switch, and a significant increase in the number of video users on some APs compared to other APs, along with an increase in the wireless retransmission rate and media stuttering rate of the corresponding APs. If the spatiotemporal graph neural network predicts that the probability of high load and access link congestion for a specific AP exceeds the threshold in the near future, the system can implement strategies such as pre-caching of popular content, prioritizing video streams, restricting access to high-load APs, guiding users to low-load APs, and delaying non-real-time background synchronization, thereby proactively releasing resources before severe congestion occurs. When a large number of passengers are concentrated in a certain cabin segment using mobile terminals to access the network, uneven load may occur between adjacent access points (APs). The system constructs a dynamic graph based on the number of APs accessed, wireless signal quality, user service type, and the status of adjacent APs. When predicting that the load of a certain AP will increase rapidly, the system can adjust the access guidance parameters of adjacent APs, divert new users, and adjust the channel or power when the equipment capacity allows. For established media playback connections, the system prioritizes maintaining connection stability to avoid playback interruptions caused by large-scale forced switching. When multiple SDUs need to update or synchronize content, if the synchronization task and passenger video service occupy the access link at the same time, it may cause media playback to stutter. When the system predicts that the synchronization traffic of a specified SDU group is large, the video service on the same link is increasing, and continuing synchronization will increase the risk of congestion, the SDUs can be divided into multiple synchronization groups, and the synchronization of non-real-time content can be temporarily suspended. Synchronization can be resumed after the link load decreases. Limited bandwidth is allocated to synchronization tasks that must be completed immediately to ensure that the media service being played by passengers and the guarantee service are not significantly affected. To verify the effectiveness of the proposed method, an airborne cabin network experimental platform or a digital twin simulation environment can be constructed to compare the fixed configuration scheme, the threshold triggering scheme, and the proposed ST-GNN active prediction scheme, as shown in Table 3: Table 3 Test Scenario Design

[0017] Evaluation metrics include average number of media stutters, first frame loading time, maximum AP load rate, load variance of each AP, peak utilization of backbone links, access link congestion duration, cache hit rate, minimum bandwidth satisfaction rate for guaranteed services, number of policy adjustments, and system anomaly recovery time. Among these, a lower AP load variance indicates a more balanced wireless access load; a shorter access link congestion duration indicates a more timely resource allocation policy; and the minimum bandwidth satisfaction rate for guaranteed services should be maintained above the preset requirements. The evaluation metrics are shown in Table 4. Table 4 Key Evaluation Indicators

[0018] This invention addresses the challenges of high-concurrency access, media service bursts, and multi-device resource coupling in airborne cabin WiFi networks. It proposes a dynamic resource allocation method based on a spatiotemporal graph neural network. This method represents the heterogeneous airborne network, composed of MSUs, switches, ACs, APs, SDUs, and passenger terminals, as a dynamic graph structure. It learns the spatial relationships between different devices and links through a graph attention mechanism, extracts the changing trends of load and service requests through a time-series model, and uses a multi-task output structure to predict link congestion, AP load, QoS risks, and content requirements. Based on these predictions, the system generates resource allocation strategies and completes closed-loop optimization under constraints of service protection, configuration stability, and a security fallback mechanism. Compared to fixed configuration or threshold-triggered methods, this method can identify risks in advance before congestion and significant lag occur, and transform the prediction results into joint control actions of switches, APs, MSUs, and SDUs. It is suitable for deployment in MSUs or airborne edge computing devices to perform local intelligent scheduling. Further research can further verify the model's lightweight design, inference latency, policy security, and engineering verifiability by combining actual airborne device interfaces, the number of APs, SDU cascade structures, and concurrent test data of 300-600 users.

[0019] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for configuring airborne WiFi network resources based on spatiotemporal graph neural networks, characterized in that: Includes the following steps: S1: Construct a heterogeneous graph of the airborne cabin WiFi network, representing the network state as a heterogeneous graph containing multiple types of device nodes and various connection edges; S2: Collect network states within a continuous time window to form a dynamic graph sequence, input it into a spatiotemporal graph neural network model, and predict network resource requirements and risk status at future moments; S3: Based on the prediction results of S2, generate a set of candidate resource configuration strategies that include multiple types of device control actions; S4: Perform a comprehensive evaluation on each candidate strategy generated in S3, and select the strategy that satisfies the constraints and has the smallest evaluation function value as the target resource allocation strategy. S5: Send the target resource configuration strategy selected in S4 to the corresponding device for execution, and make feedback and corrections based on the execution results.

2. The airborne WiFi network resource allocation method based on spatiotemporal graph neural network according to claim 1, characterized in that: The S1 heterogeneous graph includes a set of nodes, a set of edges, node features, and edge features; The expression for the heterogeneous graph is as follows: ; in, Represents a set of device nodes. Represents the set of connecting edges. Represents the node state feature matrix. Represents the edge state feature matrix; The node set consists of multiple types of device nodes and user service nodes, and the expression of the node set is as follows: ; in, This indicates the airborne media service unit node. Indicates a switch node. Indicates a wireless controller node. Indicates a wireless access point node. Indicates the terminal unit node. This indicates a passenger mobile terminal node or a user aggregation node; The expression for the set of connecting edges is as follows: ; in, This indicates the backbone link between the MSU and the switch. This indicates the access link between the switch and the AC, AP, or SDU. This indicates the wireless connection between the AP and the passenger terminal. This indicates the logical business relationship for distributing media content from the MSU to users or terminals. This indicates a cascaded link or backup link between SDUs; the expression for the node characteristics is as follows: ; in, This means that for any node At any moment The node feature vectors, For node type, For processor utilization, For memory usage, For business traffic, For the length of the request queue, For the number of connected users, In cached state, The device is either online or in an abnormal state; The expression for the edge feature is as follows: ; in, Represents any connecting edge The edge feature vectors, Indicates the link type. Indicates the nominal bandwidth. Indicates real-time utilization rate. Indicates transmission delay. Indicates packet loss rate. Indicates latency jitter. This indicates bit errors or abnormal latency jitter. This indicates the error or anomaly count.

3. The airborne WiFi network resource allocation method based on spatiotemporal graph neural network according to claim 2, characterized in that: The expression for the S2 dynamic graph sequence is as follows: ; in, Indicates the current scheduling time A dynamic graph sequence consisting of network state graphs at L consecutive sampling times, with the endpoint being the endpoint. , and These represent the heterogeneous network state diagrams corresponding to the start time t-L+1, the next sampling time t-L+2, and the current time t, respectively, where L represents the number of historical sampling steps included in the input time window. The S2 prediction of future network resource demands and risk states is based on the spatiotemporal graph neural network model given... Predicting the future under certain conditions The network resource demand and risk status within each time step are expressed as follows: ; in, This represents the joint multi-task prediction results from time t+1 to t+H, including link congestion probability, AP load, QoS risk, and content requirements. H represents the number of future time steps included in the prediction time span. This represents a spatiotemporal graph neural network mapping function that takes a dynamic graph sequence as input and outputs the future network resource demand and risk status. The spatiotemporal graph neural network model consists of a heterogeneous feature encoding layer, a spatial graph attention layer, a time series feature extraction layer, a multi-task prediction output layer, and a confidence evaluation layer. The relevant expression for the heterogeneous feature encoding layer is as follows: ; in, Represents a node At any moment The initial hidden features, This represents the encoding network corresponding to the node type. This means that for any node At any moment Node feature vectors; The relevant expression for the spatial graph attention layer is as follows: ; in, Represents a node No. Spatial features after layer update It is a non-linear activation function. and For a trainable parameter matrix, Represents a node No. Spatial characteristics of layers Represents a node No. Spatial characteristics of layers For nodes For nodes Attention weights Represents a set of edge types. Indicates passing through the first Edges and Nodes The set of connected neighbor nodes, and This represents the summation operation; The relevant expressions for the time series feature extraction layer are as follows: ; in, Represents a node The spatiotemporal fusion representation at the current scheduling time t, This represents a gated recurrent unit network used to extract the temporal dependencies of node states. , and These represent the spatial features of node v at the beginning time t-L+1, subsequent time t-L+2, and current time t of the input time window, respectively, after passing through the spatial graph attention layer, where L represents the length of the time window. The multi-task prediction output layer is used to simultaneously predict link congestion, AP load, QoS risk, and content demand. The prediction expression for link congestion is as follows: ; in, Indicates connecting edges The probability of congestion occurring within a future predicted time period. and Let i and j represent the spatiotemporal fusion representations of the two endpoints of the edge at the current time t, respectively. This represents the edge feature vector connecting edge (i,j) at the current time t. This represents a link prediction decoder that maps node representations and edge features to link congestion probabilities. The prediction expression for the AP load is as follows: ; in, This indicates the target AP's load rate or the expected number of connected users over a future forecast period. This represents the spatiotemporal fusion representation of the target AP at the current time t. This refers to an AP load prediction decoder that maps the spatiotemporal fusion representation of APs to future load rates or the number of connected users. The relevant expression for the QOS risk is as follows: ; in, This indicates the probability of media playback stuttering, increased first-frame waiting time, or decreased service quality. This represents the spatiotemporal fusion representation of the target media service or service aggregation node at the current time t. This represents a QoS risk prediction decoder that maps the spatiotemporal fusion representation of business to the probability of service quality degradation. The relevant expressions for the content requirements are as follows: ; in, This indicates the expected access demand for the specified content within a future time period. This represents the spatiotemporal fusion representation of the airborne media service unit node at the current time t. This represents the spatiotemporal fusion representation of the target media service or user service aggregation node related to the specified content at the current time t. This refers to a content demand prediction decoder that predicts future access needs for specified content based on MSU status and business demand characteristics. The confidence evaluation layer is used to improve the accuracy of link load prediction, AP load prediction, QoS risk classification, and content demand prediction. The relevant expressions are as follows: ; in, This represents the value of the comprehensive loss function; To account for errors in link utilization and congestion risk prediction, Errors in predicting the number of AP users or load rate. The classification error is due to a decline in the quality of media services. For content demand prediction error, To ensure compliance with penalties for violations of business resource constraints, , , , and These are the weighting coefficients for each loss term.

4. The airborne WiFi network resource allocation method based on spatiotemporal graph neural network according to claim 3, characterized in that: The expression for the S3 candidate resource configuration strategy set is as follows: ; in, This represents the set of candidate resource allocation strategies generated at time t. , and These represent the 1st, 2nd, and nth candidate resource configuration policies in the set, respectively, where n represents the total number of candidate policies. Each candidate policy consists of one or more control actions from the switch, AP / AC, MSU, and SDU.

5. The airborne WiFi network resource allocation method based on spatiotemporal graph neural network according to claim 4, characterized in that: The expression for the S4 comprehensive evaluation is as follows: ; in, This represents the evaluation function value. Indicate candidate strategies, This represents the predicted link congestion cost after implementing the strategy. This indicates the predicted cost of media buffering. This indicates the cost of uneven load distribution among the access points (APs). This indicates the stability cost corresponding to the adjustment range and frequency. This indicates the risk and cost of impacting service security or causing link unavailability. , , , and To evaluate the cost of the indicators, , , , and As part of the evaluation indicators, This indicates the actual bandwidth obtained by the guaranteed service. This indicates the minimum bandwidth threshold required to guarantee service availability. The expression for selecting the strategy that satisfies the constraints and minimizes the evaluation function value as the target resource allocation strategy is as follows: ; in, This represents the objective resource allocation strategy that satisfies all constraints and minimizes the evaluation function value. This represents the set of candidate resource allocation strategies at time t. This represents the minimization selection operation, which searches for and returns the strategy that minimizes the evaluation function value from the set of candidate resource allocation strategies.