MU-MIMO communication method and system of airborne Wi-Fi system based on hybrid expert model MoE

Abstract

The invention relates to an MU-MIMO communication method and system of an airborne Wi-Fi system based on a hybrid expert model MoE, and the method comprises the steps: taking a CSI sequence as a time sequence input, carrying out the multi-scale channel dynamic modeling through a Transform network integrated with an MoE module, and improving the calculation efficiency through expert sparse activation; the user scheduling and cluster selection module outputs a plurality of alternative space scheduling strategies by using MoE, and performs expert selection in combination with a historical success rate and an environment state; the MIMO precoding parameter generation module routes to a corresponding precoding strategy sub-network according to an input state characteristic based on an MoE expert routing mechanism; all modules are integrated with uncertainty estimation, and when MoE output distribution entropy is high or expert conflicts are serious, a traditional channel estimation and scheduling strategy is automatically switched to. According to the method, the channel estimation precision, the scheduling efficiency and the transmission performance of the MU-MIMO system in the complex environment of the aircraft cabin can be remarkably improved, and meanwhile, the reliability and the real-time performance of the system are ensured.

Description

Technical Field

[0001] This invention relates to the field of communication technology, and specifically to a MU-MIMO communication method and system for an airborne Wi-Fi system based on a hybrid expert model (MoE). Background Technology

[0002] With the widespread adoption of Wi-Fi services in aircraft cabins, multi-user MIMO (MU-MIMO) technology has become crucial for enhancing system capacity and user experience during in-cabin deployments. MU-MIMO systems effectively utilize spatial resources to increase throughput by transmitting signals to multiple users simultaneously in the same time slot. However, the complex multipath channel environment within aircraft cabins, coupled with dynamic user distribution and variable interference, presents challenges for channel estimation, user scheduling, and precoding. Traditional MIMO channel estimation and precoding methods (such as ZF and WMMSE) assume ideal channels or Gaussian noise, making it difficult to adapt quickly to non-stationary, reflection, and interference environments. Simultaneously, statistical or rule-based user scheduling methods (such as polling or channel quality allocation) cannot fully capture complex dynamic characteristics, leading to decreased spectrum utilization efficiency.

[0003] In recent years, deep learning and large-scale model technology have attracted attention in the field of wireless communication. Models such as Transformer have been used for wireless channel modeling and feature prediction. For example, a Transformer-based method for UAV air-to-ground channel prediction has been proposed, and its superior performance in modeling complex dynamic channels has been verified. Large-scale pre-trained models can learn complex channel environment features and be applied in MIMO systems to achieve near-optimal precoding and scheduling. On the other hand, the Mixture-of-Experts (MoE) architecture, as a sparse expert mechanism, can reduce computational costs by activating only a subset of experts through gating networks while maintaining large model capacity. The expert splitting and sparse activation mechanism of MoE has demonstrated efficient scalability in fields such as natural language processing, and each expert can focus on processing specific types of input patterns. Introducing the MoE concept into communication systems is expected to enable specialized processing strategies for different channel or spatial groups while taking into account resource utilization.

[0004] Furthermore, for mission-critical scenarios such as cabin communication, the reliability requirements of the system are extremely high. Simply relying on the output of neural networks carries the risk of uncertainty; when the model encounters an "unseen" environment or expert routing conflicts occur, performance may become unpredictable. Therefore, combining uncertainty estimation mechanisms with traditional channel processing methods to form a dynamic backup strategy is of great significance. On the other hand, the resources of aviation edge computing platforms are limited, thus requiring optimization of large models, such as using expert sparse activation and model pruning to reduce model parameters and inference latency. In summary, there is currently a lack of technical solutions for systematically applying large models and MoE structures to aircraft cabin MU-MIMO channel modeling, user scheduling, and precoding. Summary of the Invention

[0005] The purpose of this invention is to overcome the problems existing in the prior art and provide a MU-MIMO communication method and system for airborne Wi-Fi systems based on the hybrid expert model MoE, which solves the problems of low efficiency in channel modeling, user scheduling and precoding, large impact of performance on environmental changes and poor system reliability.

[0006] To achieve the above-mentioned technical objectives and effects, the present invention is implemented through the following technical solution: A MU-MIMO communication method for an airborne Wi-Fi system based on a hybrid expert model (MoE) includes the following steps: Step S1: Obtain the CSI sequence of the multi-user MIMO system in multiple consecutive time slots, and perform normalization, interpolation, or dimensionality reduction preprocessing on the CSI sequence to form a time series input; Step S2: Feed the timing input into the Transformer network containing the MoE layer to obtain the estimation or prediction results of the current or next time slot channel; Step S3: Based on the channel characteristics and network status information in step S2, the user set and / or cooperative user group in the current time slot are determined using a user scheduling and cluster selection module containing the MoE structure. Step S4: Based on the user set, group information and channel estimation results, the downlink precoding matrix or its parameters are output using the MIMO precoding parameter generation module containing the MoE structure; Step S5: Perform uncertainty assessment on the outputs of steps S2 to S4. When the uncertainty exceeds the threshold, trigger the backoff mechanism to adopt the preset traditional channel processing, scheduling or precoding method.

[0007] Furthermore, in step S3, a gated router is used to calculate the fitness probability of each strategy based on the input features, generate a weight distribution, and finally select the expert strategy with the largest weight as the scheduling scheme for the current time slot.

[0008] Furthermore, the gated router operates as follows: Let the current input feature vector be x, and the router vector parameter be r. i Then calculate and normalize the activation weights for each expert: And only retain the K highest non-zero weights; the above w i The output of the MoE layer is obtained by multiplying each output by the corresponding expert output and then summing the results.

[0009] Furthermore, in step S3, the user scheduling and cluster selection module selection includes a user scheduling module using another integrated MoE structure, where different experts in the user scheduling module correspond to different spatial scheduling strategies; in each time slot, based on the scheduling success rate of the previous time slot and the current environmental state, the gating network is input to generate the scheduling probability distribution corresponding to each expert; the expert with the highest probability is selected to determine the user scheduling scheme for the current time slot, and the historical success rate information is updated.

[0010] Furthermore, in step S4, the MIMO precoding parameter generation module is designed based on the MoE expert routing mechanism. The input features include channel estimation results and selected user cluster information. The gated router calculates the routing weights of each expert according to the input features and routes the inputs to several precoding expert subnetworks. Each expert subnetwork outputs a corresponding precoding matrix or parameters. The final precoding result is a matrix synthesized by weighting the outputs of each expert according to the routing weights.

[0011] Furthermore, in step S5, each module integrates an uncertainty estimation mechanism, using the probability distribution entropy output by MoE as one of the uncertainty measures; when the output distribution entropy of MoE exceeds a preset threshold or when multiple expert output results conflict severely, a backoff switching mechanism is automatically triggered to replace the corresponding module with traditional channel estimation, user scheduling or precoding methods to ensure communication reliability.

[0012] Furthermore, the channel estimation obtains features at different time scales through multi-scale convolution or multi-head self-attention, and the final fusion strategy is determined by the MoE layer; the MoE expert of the user scheduling module learns the spatial channel correlation between different users to correspond to different cooperative coverage groups; the MoE expert of the MIMO precoding parameter generation module designs exclusive precoding schemes for different spatial groups by jointly optimizing multi-user channel information.

[0013] Furthermore, the user scheduling and cluster selection module selects policy candidates based on the output of the gating network, determines the success rate of each policy in different environments by accumulating historical data, and adjusts the expert selection weights based on Bayesian or deep reinforcement learning methods to dynamically adapt to changes in spatial location and interference distribution.

[0014] Furthermore, the precoding expert subnetwork adopts a linear ZF precoding, WMMSE precoding, or deep neural network structure, which is automatically selected and switched by the gated router according to the input channel characteristics, so as to balance precoding efficiency and accuracy; the output matrices of each expert are linearly combined in the frequency domain to achieve the final multi-user downlink precoding.

[0015] A MU-MIMO communication system for an airborne Wi-Fi system based on a hybrid expert model (MoE) is disclosed. The MU-MIMO communication system includes processing modules connected by a high-speed internal bus, a hardware acceleration unit with sparse computation acceleration function, and a memory. The processing modules include a channel estimation and modeling module, a user scheduling and cluster selection module, a MIMO precoding parameter generation module, an uncertainty estimation and switching module, and an edge deployment management module.

[0016] The beneficial effects of this invention are: This invention introduces the Transformer large model and the Mixture-of-Experts architecture, achieving a leapfrog optimization of channel modeling and resource scheduling in the complex MU-MIMO communication environment of an aircraft cabin: Efficiency Improvement: This invention utilizes the sparse activation mechanism of MoE to significantly expand the model capacity without increasing the computational overhead proportionally. Combined with gated routing and multi-expert parallel strategies, the system can automatically select the optimal strategy in different scenarios, thereby improving resource utilization.

[0017] Performance enhancements: The Transformer network provides powerful temporal feature extraction capabilities and can model dynamic cabin channels; the MoE structure allows each expert to specialize in a specific channel environment or user combination, resulting in a better overall approximation of the optimal precoding and scheduling scheme.

[0018] Good robustness: The integrated uncertainty estimation and switching mechanism enables the system to actively revert to traditional algorithms when faced with model uncertainty or sudden environmental changes, ensuring that the system performance will not drop drastically due to misjudgment by the deep model.

[0019] Edge feasibility: By reducing model complexity through expert pruning and sparse activation, large models can achieve real-time inference on airborne edge platforms. This solution balances high performance with hardware resource constraints and is an innovative technology suitable for aviation communications. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall framework and data flow of the present invention; Figure 2 This is a diagram showing the internal structure of the channel modeling module of the present invention; Figure 3 This is a flowchart of the user scheduling and collaborative cluster module of the present invention; Figure 4 This is a schematic diagram of the MIMO precoding module of the present invention; Figure 5 This is a schematic diagram of the uncertainty estimation and switching mechanism of the present invention. Detailed Implementation

[0021] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0022] A MU-MIMO communication method for an airborne Wi-Fi system based on a hybrid expert model (MoE) includes the following steps: Step S1: Obtain the CSI sequence of the multi-user MIMO system in multiple consecutive time slots, and perform normalization, interpolation, or dimensionality reduction preprocessing on the CSI sequence to form a time series input; Step S2: The timing input is fed into the Transformer network containing the MoE layer to obtain the estimation or prediction results of the current or next time slot channel. This multi-layer Transformer-MoE structure model is deployed on the aircraft's airborne edge computing platform and executed using professional acceleration hardware. By implementing techniques such as expert sparse activation, model pruning, and quantization, the number of model parameters and inference computation complexity are significantly reduced, thereby improving real-time performance. Step S3: Based on the channel characteristics and network status information in step S2, the user set and / or cooperative user group in the current time slot are determined using a user scheduling and cluster selection module containing the MoE structure. Step S4: Based on the user set, group information and channel estimation results, the downlink precoding matrix or its parameters are output using the MIMO precoding parameter generation module containing the MoE structure; Step S5: Perform uncertainty assessment on the outputs of steps S2 to S4. When the uncertainty exceeds the threshold, trigger the backoff mechanism to adopt the preset traditional channel processing, scheduling or precoding method.

[0023] In step S3, a gated router is used to calculate the fitness probability of each strategy based on the input features, generate a weight distribution, and finally select the expert strategy with the largest weight as the scheduling scheme for the current time slot.

[0024] The gated router operates as follows: Let the current input feature vector be x, and the router vector parameter be r. i Then calculate and normalize the activation weights for each expert: And only retain the K highest non-zero weights; the above w i The output of the MoE layer is obtained by multiplying each output by the corresponding expert output and then summing the results.

[0025] In step S3, the user scheduling and cluster selection module selection includes using another integrated MoE structure user scheduling module, where different experts in the user scheduling module correspond to different spatial scheduling strategies; in each time slot, based on the scheduling success rate of the previous time slot and the current environmental state, the gating network is input to generate the scheduling probability distribution corresponding to each expert; the expert with the highest probability is selected to determine the user scheduling scheme for the current time slot, and the historical success rate information is updated.

[0026] In step S4, the MIMO precoding parameter generation module is designed based on the MoE expert routing mechanism. The input features include channel estimation results and selected user cluster information. The gated router calculates the routing weights of each expert according to the input features and routes the inputs to several precoding expert subnetworks. Each expert subnetwork outputs a corresponding precoding matrix or parameters. The final precoding result is a matrix synthesized by weighting the outputs of each expert according to the routing weights.

[0027] In step S5, each module integrates an uncertainty estimation mechanism, using the probability distribution entropy output by the MoE as one of the uncertainty measures. When the distribution entropy output by the MoE exceeds a preset threshold or when multiple expert outputs conflict severely, a fallback switching mechanism is automatically triggered to replace the corresponding module with traditional channel estimation, user scheduling, or precoding methods to ensure communication reliability. In addition to the distribution entropy criterion, the uncertainty estimation mechanism can also combine the variance output by the integrated model or use methods such as Monte Carlo Dropout to quantify the output confidence. If the confidence is lower than a preset standard, a fallback mechanism is triggered.

[0028] The channel estimation obtains features at different time scales through multi-scale convolution or multi-head self-attention, and the final fusion strategy is determined by the MoE layer; the MoE experts of the user scheduling module learn the spatial channel correlation between different users to correspond to different cooperative coverage groups; the MoE experts of the MIMO precoding parameter generation module design exclusive precoding schemes for different spatial groups by jointly optimizing multi-user channel information.

[0029] The user scheduling and cluster selection module selects policy candidates based on the output of the gating network, determines the success rate of each policy in different environments by accumulating historical data, and adjusts the expert selection weights based on Bayesian or deep reinforcement learning methods to dynamically adapt to changes in spatial location and interference distribution.

[0030] The precoding expert subnetwork adopts linear ZF precoding, WMMSE precoding, or deep neural network structure, which is automatically selected and switched by the gated router according to the input channel characteristics, so as to balance precoding efficiency and accuracy; the output matrices of each expert are linearly combined in the frequency domain to achieve the final multi-user downlink precoding.

[0031] A MU-MIMO communication system based on a hybrid expert model (MoE) for airborne Wi-Fi systems, such as Figure 1 As shown, the MU-MIMO communication system includes processing modules connected by a high-speed internal bus, a hardware acceleration unit with sparse computation acceleration capabilities, and a memory. The modules work collaboratively through hardware facilities such as controllers and memory. The processing modules include a channel estimation and modeling module, a user scheduling and cluster selection module, a MIMO precoding parameter generation module, an uncertainty estimation and handover module, and an edge deployment management module. Furthermore, the MU-MIMO communication system also includes a database storing historical success rate data and expert performance metrics, as well as an edge training module for updating model parameters online or offline, enabling the MoE model to continuously learn new channel and scheduling environment characteristics. The following provides a detailed description of each module: Channel estimation and modeling module: such as Figure 2 As shown, the collected multi-user MIMO feedback CSI sequences are treated as time-series data, and multi-scale channel features in the time and frequency domains are extracted through a multi-layer Transformer network. Each Transformer network contains a self-attention layer and a feedforward layer, where the feedforward layer is replaced by a hybrid expert MoE layer. Each MoE layer consists of several expert sub-networks and a gated network Router. Given an input feature vector x, the gated network first calculates its interaction with the routing vectors r of each expert. i The similarity (e.g., inner product) is then normalized using Softmax to obtain the expert activation weight w. i Only the Top-K experts are selected for calculation; specifically, it can be expressed as: , , In the formula, f i (·) is the mapping function of the i-th expert network; this gating mechanism ensures that only a few experts are activated in each layer, so that the computational load does not increase while the model parameter scale increases significantly; the superposition of Transformer layers of multiple scales can dynamically capture the time-varying channel characteristics caused by factors such as Doppler effect and multipath reflection caused by aircraft cabin movement; User scheduling and cluster selection module: such as Figure 3As shown, during the transmission scheduling phase, another MoE-based network is set up to generate optional user scheduling strategies. The input to this module includes user channel characteristics extracted by the channel modeling module and current network state information (such as interference intensity, user service requirements, etc.). This module also adopts a multi-expert structure, with different experts corresponding to different spatial scheduling strategies or user group partitions. For example, one expert may tend to select user groups with low channel correlation, while another expert may prioritize scheduling users with low historical throughput. The gated router calculates the fitness probability of each strategy based on the input characteristics and generates a weight distribution. Finally, the expert strategy with the largest weight is selected as the scheduling scheme for the current time slot. In addition, the system continuously records the success rate of each strategy under different environments and uses it as feedback for expert selection at the next moment to achieve dynamic optimization. MIMO precoding parameter generation module: such as Figure 4 As shown, this module generates precoding based on the MoE expert routing mechanism. The input to this module includes the spatial channel matrix after channel estimation and user cluster information. Several expert sub-networks are set up, each expert corresponding to a precoding algorithm (such as linear ZF, WMMSE, or deep precoding networks). A gated router calculates expert weights based on the input channel feature vectors and routes the input to the corresponding expert. Each expert sub-network can internally be a small deep network or a parameterized function used to generate the precoding matrix for the user cluster. The final precoding matrix is ​​obtained by linearly combining the outputs of each expert according to the routing weights. In this way, the system can flexibly select the most suitable precoding strategy according to different channel environments and scheduling schemes, thus achieving a balance between accuracy and complexity. Uncertainty estimation and switching module: such as Figure 5 As shown, to ensure system reliability, output uncertainty estimates are integrated into each of the above modules; after each MoE output, the entropy of the expert weight distribution is calculated: Furthermore, it can combine the output variance obtained, such as Monte Carlo Dropout, to assess the confidence of the current decision. When the entropy exceeds the threshold or there is a significant conflict in expert routing (multiple experts with similar weights), the system triggers a backup mechanism: temporarily using traditional channel estimation, classical scheduling algorithms (such as CSI-based greedy scheduling), or linear precoding methods until the deep model returns a sufficiently confident output. This method combines the advantages of deep models and traditional methods, improving the reliability and robustness of the communication process. Edge Deployment Management Module: The aforementioned Transformer-MoE model is deployed on an aircraft's airborne edge computing platform, supporting parallel signal processing tasks. To reduce model computation and inference latency, expert sparse activation and pruning optimization are performed on the model. Specifically, for each MoE layer, only the calculations of the Top-K experts are performed, while unselected experts remain idle. After model training, pruning techniques are used to remove expert subnetworks that have been inactive for a long time or have low contribution, and the weights can be quantized. Related research shows that neural network pruning for edge devices can significantly improve inference speed and energy efficiency with almost no loss of accuracy. In summary, the design of this invention balances model performance and resource constraints, making it suitable for real-time communication processing in harsh environments such as airborne environments. Finally, there is the framework integration: combining the above functional modules into a complete system; the system can automatically adjust its operating strategy according to the flight phases such as takeoff and landing, cruise, and the distribution of users in the cabin, ensuring efficient communication around the clock. To illustrate the above technical solution, a partial mathematical description can be provided; for example, let H be the CSI matrix collected at time t. t After preprocessing, it forms a time-series input sequence x. t The output of a feedforward layer in the Transformer-MoE model can be represented as: , In the formula, f i (·) represents the mapping function of the i-th expert network; r i For the corresponding routing vector, calculation is performed only for the top K experts by score; the output entropy is defined as: When H(x) t Exceeding threshold H th Alternatively, if the difference between the largest and second-largest weights is less than a threshold, a backup strategy is triggered; in terms of user scheduling, the historical success rate vector m and the current environment state s can be used. t Input the data into a gating network and calculate the estimated revenue q for each expert. i And select the scheduling scheme corresponding to the largest one.

[0032] The following embodiment further illustrates the technical solution of the present invention in detail, but the embodiments of the present invention are not limited thereto. Example

[0033] A MU-MIMO communication system based on the method of this invention is deployed in an aircraft cabin. The system uses an 8×8 antenna array to provide concurrent downlink transmission for multiple passenger terminals. User terminals feed back downlink CSIs at a preset period, forming a CSI sequence of length T; the CSI sequence is normalized and dimensionality reduced, and then used as input to a Transformer-MoE model. In this embodiment, the model contains 3 Transformer modules, each layer has E=32 experts and uses Top-K=4 sparse activation.

[0034] After completing channel estimation, the scheduling module sets up three scheduling experts, corresponding to: (a) a scheduling strategy biased towards reducing spatial correlation between users; (b) a scheduling strategy biased towards fairness; and (c) a scheduling strategy biased towards maximizing instantaneous throughput. The gated router outputs a weight distribution based on the current channel characteristics and historical throughput indicators, and selects the strategy with the highest weight to determine the user set and cooperative group for the current slot.

[0035] The precoding module employs four precoding experts: linear ZF, WMMSE, and two parameterized deep precoding networks. The gated router selects K experts based on the input channel characteristics to participate in the computation, and then weights and fuses the outputs of each expert to obtain the precoding matrix.

[0036] In each stage of channel modeling, scheduling, and precoding, the system calculates uncertainty indicators such as gating weight distribution entropy and weight difference. When the uncertainty exceeds a threshold, the system can bypass and switch to traditional algorithms (such as CSI-based greedy scheduling, PF scheduling, or ZF / WMMSE precoding), and record the environmental state and execution results of that time slot for subsequent model updates.

[0037] Other implementation methods may include: adjusting the number of experts and the Top-K value, introducing uplink / downlink joint modeling, or adding specific engineering constraints (such as power limits, rate thresholds, etc.) to the fallback strategy. Any equivalent substitutions or modifications made without departing from the inventive concept should fall within the protection scope of this invention.

[0038] This invention provides an innovative solution to the channel modeling, user scheduling, and precoding problems in MU-MIMO communication scenarios of airborne Wi-Fi systems by integrating the MoE architecture and Transformer large model technology, and has significant performance and engineering implementation value.

[0039] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A MU-MIMO communication method for an airborne Wi-Fi system based on a hybrid expert model (MoE), characterized in that, The method includes the following steps: Step S1: Obtain the CSI sequence of the multi-user MIMO system in multiple consecutive time slots, and perform normalization, interpolation, or dimensionality reduction preprocessing on the CSI sequence to form a time series input; Step S2: Feed the timing input into the Transformer network containing the MoE layer to obtain the estimation or prediction results of the current or next time slot channel; Step S3: Based on the channel characteristics and network status information in step S2, the user set and / or cooperative user group in the current time slot are determined using a user scheduling and cluster selection module containing the MoE structure. Step S4: Based on the user set, group information and channel estimation results, the downlink precoding matrix or its parameters are output using the MIMO precoding parameter generation module containing the MoE structure; Step S5: Perform uncertainty assessment on the outputs of steps S2 to S4. When the uncertainty exceeds the threshold, trigger the backoff mechanism to adopt the preset traditional channel processing, scheduling or precoding method.

2. The MU-MIMO communication method for an airborne Wi-Fi system based on a hybrid expert model (MoE) according to claim 1, characterized in that, In step S3, a gated router is used to calculate the fitness probability of each strategy based on the input features, generate a weight distribution, and finally select the expert strategy with the largest weight as the scheduling scheme for the current time slot.

3. The MU-MIMO communication method for an airborne Wi-Fi system based on a hybrid expert model (MoE) according to claim 2, characterized in that, The gated router operates as follows: Let the current input feature vector be x, and the router vector parameter be r. i Then calculate and normalize the activation weights for each expert: And only retain the K highest non-zero weights; the above w i The output of the MoE layer is obtained by multiplying each output by the corresponding expert output and then summing the results.

4. The MU-MIMO communication method for an airborne Wi-Fi system based on a hybrid expert model (MoE) according to claim 1, characterized in that, In step S3, the user scheduling and cluster selection module selection includes a user scheduling module using another integrated MoE structure. Different experts in the user scheduling module correspond to different spatial scheduling strategies. In each time slot, the scheduling success rate of the previous time slot and the current environmental state are input into the gating network to generate the scheduling probability distribution corresponding to each expert. The expert with the highest probability is selected to determine the user scheduling scheme for the current time slot, and the historical success rate information is updated.

5. The MU-MIMO communication method for an airborne Wi-Fi system based on a hybrid expert model (MoE) according to claim 2, characterized in that, In step S4, the MIMO precoding parameter generation module is designed based on the MoE expert routing mechanism. The input features include channel estimation results and selected user cluster information. The gated router calculates the routing weights of each expert according to the input features and routes the inputs to several precoding expert subnetworks. Each expert subnetwork outputs a corresponding precoding matrix or parameters. The final precoding result is a matrix synthesized by weighting the outputs of each expert according to the routing weights.

6. The MU-MIMO communication method for an airborne Wi-Fi system based on a hybrid expert model (MoE) according to claim 1, characterized in that, In step S5, each module integrates an uncertainty estimation mechanism, using the probability distribution entropy of the MoE output as one of the uncertainty measures. When the MoE output distribution entropy exceeds a preset threshold or when multiple expert output results conflict severely, a backoff switching mechanism is automatically triggered to replace the corresponding module with traditional channel estimation, user scheduling, or precoding methods to ensure communication reliability.

7. The MU-MIMO communication method for an airborne Wi-Fi system based on a hybrid expert model (MoE) according to claim 5, characterized in that, The channel estimation obtains features at different time scales through multi-scale convolution or multi-head self-attention, and the final fusion strategy is determined by the MoE layer; the MoE experts of the user scheduling module learn the spatial channel correlation between different users to correspond to different cooperative coverage groups; the MoE experts of the MIMO precoding parameter generation module design exclusive precoding schemes for different spatial groups by jointly optimizing multi-user channel information.

8. The MU-MIMO communication method for an airborne Wi-Fi system based on a hybrid expert model (MoE) according to claim 4, characterized in that, The user scheduling and cluster selection module selects policy candidates based on the output of the gating network, determines the success rate of each policy in different environments by accumulating historical data, and adjusts the expert selection weights based on Bayesian or deep reinforcement learning methods to dynamically adapt to changes in spatial location and interference distribution.

9. The MU-MIMO communication method for an airborne Wi-Fi system based on a hybrid expert model (MoE) according to claim 7, characterized in that, The precoding expert subnetwork adopts linear ZF precoding, WMMSE precoding, or deep neural network structure, which is automatically selected and switched by the gated router according to the input channel characteristics, so as to balance precoding efficiency and accuracy; the output matrices of each expert are linearly combined in the frequency domain to achieve the final multi-user downlink precoding.

10. A MU-MIMO communication system according to any one of claims 1 to 9, characterized in that, The MU-MIMO communication system includes processing modules connected by a high-speed internal bus, a hardware acceleration unit with sparse computing acceleration function, and a memory. The processing modules include a channel estimation and modeling module, a user scheduling and cluster selection module, a MIMO precoding parameter generation module, an uncertainty estimation and switching module, and an edge deployment management module.

Images