Multi-unmanned aerial vehicle NOMA communication de-noising method based on pre-training basic model

Abstract

The invention discloses a multi-unmanned aerial vehicle NOMA communication de-noising method based on a pre-training basic model. According to the method, through self-supervised / semi-supervised pre-training, the basic model learns network physical rules and space-time dynamic characteristics, a decoder is removed during reasoning, and an encoder is adopted to be matched with a multi-task prediction head to realize efficient parallel decision making. Specifically, multi-modal embedding and space-time diagram position coding are introduced, heterogeneous data of the unmanned aerial vehicle are fused, non-Euclidean space-time topology is captured, and a final model can simultaneously predict an NOMA resource block grouping strategy, the transmitting power and position of the unmanned aerial vehicle and a performance index under the current grouping strategy in a single step, so that traditional iterative optimization is replaced; and the real-time performance and generalization capability of joint optimization of spectrum efficiency, fairness and robustness in the dynamic unmanned aerial vehicle network are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of AI-enabled collaborative NOMA communication for drones, and particularly relates to a noise reduction method for multi-drone NOMA communication based on a pre-trained basic model. Background Technology

[0002] Unmanned Aerial Vehicle (UAV) communication networks, with their flexible deployment capabilities and mobility, have demonstrated broad potential in various complex application scenarios. To improve system spectrum efficiency (SE) and support large-scale device access, Non-Orthogonal Multiple Access (NOMA) technology has been introduced into UAV networks, enabling multiple users to communicate on the same time and frequency resources. However, the performance of this technology is highly dependent on the effective control of various types of interference. In UAV ad hoc networks, the communication link structure is complex, especially when UAVs operate in full-duplex mode. The system simultaneously faces interference from both the NOMA downlink and uplink, as well as significant self-interference issues arising from full-duplex communication. This makes interference management a key challenge in system design and optimization.

[0003] To address the aforementioned interference issues and network performance optimization needs, existing technologies primarily focus on two categories of approaches. One category is optimization methods based on traditional mathematical programming. Early research typically employed techniques such as convex optimization and alternating optimization to handle resource allocation problems. For example, this involved jointly optimizing the 3D deployment and communication resources of UAVs to serve multiple users, or designing channel and power allocation algorithms to reduce interference. However, these methods have significant limitations when dealing with highly dynamic network environments. The other category comprises deep learning and reinforcement learning methods that have emerged in recent years. To adapt to the dynamic characteristics of UAV networks and the complexity of channel information, researchers have attempted to introduce deep reinforcement learning (DRL) for interference management even without knowing the channel information, or utilize artificial intelligence algorithms such as graph neural networks (GNNs) for resource optimization, aiming to adapt to complex and ever-changing communication environments through learning mechanisms.

[0004] Nevertheless, existing technologies still have several shortcomings in addressing the practical needs of drone NOMA networks:

[0005] First, traditional mathematical programming methods are ill-suited to highly dynamic network environments. They typically suffer from high variable dimensionality and slow convergence. More seriously, they often rely on oversimplification of the system model, such as assuming ideal channel state information (CSI) or ignoring some interference factors. This makes them unsuitable for real-time decision-making in scenarios where topology and channels change rapidly. The fundamental reason is that the joint resource optimization problem in UAV NOMA networks (involving location, power, and NOMA grouping) is inherently a highly coupled non-convex optimization problem. Traditional iterative algorithms have limited efficiency, and obtaining complete channel information incurs significant signaling overhead, making them impractical in decentralized, highly mobile UAV networks.

[0006] Secondly, existing deep learning-based methods generally suffer from weak generalization ability and insufficient robustness. When the network environment changes (such as an increase or decrease in the number of drones or a change in the topology), the performance of the trained deep reinforcement learning or graph neural network models drops sharply, often requiring a large amount of resources for retraining. This is mainly because existing models only fit data for specific scenarios and fail to learn and "understand" the inherent physical laws and dynamic characteristics of the UAV NOMA system from the training data, lacking the ability to build a general decision-making mechanism.

[0007] Furthermore, existing neural network models have significant shortcomings in handling heterogeneous multimodal data and their spatiotemporal dependencies in UAV networks, such as time-varying UAV positions (kinematic information), imperfect CSI, NOMA grouping strategies, and power allocation data. UAV networks inherently possess dynamic non-Euclidean topology characteristics, and the relationships between UAV nodes (channels, interference) change dynamically with their spatiotemporal positions. Existing models lack effective spatiotemporal graph encoding and multimodal embedding mechanisms, making it difficult to accurately capture the spatiotemporal dependencies of flows within the network. This results in models failing to accurately represent the network state, leading to suboptimal decisions.

[0008] Finally, existing methods perform poorly in multi-objective collaborative optimization. Unmanned aerial vehicle (UAV) communication systems need to strike a balance between multiple objectives, such as spectral efficiency (SE), fairness, and robustness. For example, simply maximizing SE may cause some links to "starve" due to excessive interference, compromising fairness. Furthermore, existing methods often struggle to maintain good fairness and robustness when network load increases (e.g., with an increase in the number of UAVs). The fundamental reason for this is the lack of a unified pre-training and optimization framework that can deeply understand the intrinsic physical relationships between multiple objectives. This makes it difficult to extract key physical characteristics from limited labeled expert data, thus hindering the achievement of effective multi-objective trade-offs under complex and multi-constraint conditions.

[0009] In summary, existing technologies lack a universal basic model framework that can understand the inherent physical laws and dynamic spatiotemporal characteristics of UAV NOMA networks, process multimodal heterogeneous data, and be efficiently generalized to different network scales and environments. This leads to difficulties in realizing UAV cooperative NOMA communication cluster technology with high spectral efficiency, high robustness, and high fairness performance indicators. Summary of the Invention

[0010] The technical problem this invention aims to solve is how to achieve rapid, real-time joint optimization of highly coupled variables such as UAV position, power, and NOMA grouping under imperfect CSI and dynamically changing environments. To this end, this invention proposes a multi-UAV NOMA communication denoising method based on a pre-trained base model. This method can grasp the inherent physical laws and dynamic spatiotemporal characteristics of UAV NOMA networks through semi-supervised / self-supervised learning, thereby overcoming the limitations of existing technologies in real-time performance, generalization, multimodal data processing, and multi-objective balancing.

[0011] To achieve the above-mentioned objectives, the present invention specifically adopts the following technical solution:

[0012] In a first aspect, the present invention provides a multi-UAV NOMA communication noise reduction method based on a pre-trained base model, comprising the following steps:

[0013] S1. A base model containing a multimodal embedding layer, a spatiotemporal graph location encoding layer, a Transformer encoder and a Transformer decoder, and a multi-task prediction head is pre-trained on a series of self-supervised or semi-supervised tasks. This enables the base model to learn the physical laws, spatiotemporal dynamic characteristics, and complex mapping relationship between the state and communication performance of the UAV NOMA network from UAV swarm operation data. After pre-training, the Transformer decoder in the base model is removed and used as the prediction model.

[0014] S2. The heterogeneous data of the real-time acquisition of the drone swarm is used as the input of the prediction model. The heterogeneous data input is mapped to a unified high-dimensional embedding space through a multimodal embedding layer. A learnable context token is introduced to aggregate global information, thereby forming the input sequence of the spatiotemporal graph location encoding layer.

[0015] S3. In the spatiotemporal graph location encoding layer of the prediction model, the node identity encoding, spatial relationship encoding and time evolution encoding of UAV NOMA network state of the UAV are concatenated and passed through a linear projection layer to obtain the spatiotemporal graph location encoding sequence, which is added to the input sequence as the embedding vector sequence that integrates spatiotemporal topological information.

[0016] S4. After the embedded vector sequence is processed by the Transformer encoder, a global average pooling operation is performed on the feature sequence output by the Transformer encoder to aggregate the feature sequence into a high-dimensional global state vector, which is used to represent the global state of the UAV NOMA network in the current time slot.

[0017] S5. In the multi-task prediction head, the global state vector is fed into multiple prediction heads in parallel to predict the NOMA resource block grouping policy, the transmit power of each UAV, the three-dimensional position vector of each UAV, the minimum signal-to-interference-plus-noise ratio of all communication links in the UAV NOMA network under the current grouping policy, and the fairness index under the current grouping policy. Based on the predicted transmit power, the power allocation coefficient between UAVs is calculated, thereby achieving multi-objective joint optimization that maximizes spectrum efficiency, fairness and robustness.

[0018] Based on the above scheme, each step can be implemented in the following preferred manner.

[0019] As a preferred embodiment of the first aspect mentioned above, in step S2, for each time slot, heterogeneous data of three modes are acquired, namely channel state information, UAV kinematic state information, and NOMA grouping strategy; the heterogeneous data of each mode is flattened into a one-dimensional vector and then passed through a multilayer perceptron corresponding to the mode, thereby mapping the flattened vector to a high-dimensional embedding space to obtain the corresponding mode embedding token; then a learnable parameterized token is used as the context token, and the context token is concatenated with all mode embedding tokens as the input sequence of the spatiotemporal graph location coding layer.

[0020] As a preferred embodiment of the first aspect above, in step S3, a learnable embedding vector is assigned to each drone number index among the multiple drones. Then, for a Token representing the relationship between two drone number indices, its node identity is encoded as the sum of the embedding vectors of the corresponding two drone number indices.

[0021] As a preferred embodiment of the first aspect mentioned above, in step S3, in the first... In each time slot, the position difference between the two UAVs is normalized and used as a relative position vector. The relative position vector is then input into a multilayer perceptron for encoding to obtain the spatial relationship code.

[0022] As a preferred embodiment of the first aspect above, in step S3, each time slot index... A learnable temporal embedding is assigned and then assigned to all tokens in the current time slot input sequence to obtain the temporal evolution code.

[0023] As a preferred embodiment of the first aspect mentioned above, in step S5, the grouping policy prediction head receives the global state vector and outputs a probability distribution representing the NOMA resource block grouping policy; the transmit power prediction head receives the global state vector, outputs the transmit power of each UAV, and the transmit power of all UAVs constitutes a first set; the position prediction head receives the global state vector, outputs the three-dimensional position vector of each UAV, and the three-dimensional position vectors of all UAVs constitute a second set; the minimum signal-to-interference-plus-noise ratio (SINNR) prediction head receives the global state vector and outputs the predicted minimum SINNR value for all communication links in the UAV NOMA network under the current grouping policy; the fairness prediction head receives the global state vector and outputs a scalar value representing the fairness index predicted under the current grouping policy.

[0024] As a preferred embodiment of the first aspect mentioned above, the grouping strategy prediction head, transmit power prediction head, position prediction head, minimum signal-to-interference-plus-noise ratio prediction head, and fairness prediction head are all multilayer perceptrons.

[0025] In a second aspect, the present invention provides a computer program product, including a computer program / instruction, which, when executed by a processor, can implement the multi-UAV NOMA communication noise reduction method based on a pre-trained basic model as described in any of the solutions in the first aspect above.

[0026] Thirdly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the multi-UAV NOMA communication noise reduction method based on a pre-trained basic model as described in any of the solutions of the first aspect above.

[0027] Fourthly, the present invention provides a computer electronic device, which includes a memory and a processor;

[0028] The memory is used to store computer programs;

[0029] The processor is configured to, when executing the computer program, implement the multi-UAV NOMA communication noise reduction method based on a pre-trained basic model as described in any of the solutions of the first aspect above.

[0030] Compared with the prior art, the present invention has the following advantages:

[0031] This invention achieves efficient real-time decision-making: during the inference phase, the Transformer decoder used in pre-training is removed, and a parallel multi-task prediction head based on encoder output is adopted instead. This design allows the model to simultaneously predict NOMA grouping, UAV launch power, fairness, and robustness metrics in a single forward propagation, thereby replacing the time-consuming iterative search required by traditional optimization algorithms (such as PSO) or the sequential decision-making process of reinforcement learning (such as PPO). This efficient inference architecture significantly reduces decision latency, meeting the stringent real-time resource scheduling requirements of highly dynamic, rapidly changing UAV networks.

[0032] This invention enhances the fusion and generalization capabilities of heterogeneous data: It designs a novel multi-modal embedding mechanism and spatiotemporal graph position coding (STG-PE). This mechanism can effectively process and fuse heterogeneous data in UAV networks (such as Channel State Information (CSI), UAV kinematic data / location, NOMA grouping strategies, etc.) and accurately capture the dynamic, non-Euclidean spatiotemporal topology of the network. Combined with a semi-supervised pre-training strategy, the model can learn the network's intrinsic physical laws from a small amount of expert-labeled data and a large amount of unlabeled data, thereby achieving excellent generalization ability and decision-making efficiency. Attached Figure Description

[0033] Figure 1 This is a flowchart of the steps of the method of the present invention;

[0034] Figure 2 This is a diagram of the pre-trained basic model architecture provided in this embodiment;

[0035] Figure 3 This is a schematic diagram illustrating the minimum SINR results of different methods provided in this embodiment;

[0036] Figure 4 This is a schematic diagram showing the total spectral efficiency results of different methods provided in this embodiment;

[0037] Figure 5 This is a schematic diagram showing the average spectral efficiency results of different methods provided in this embodiment;

[0038] Figure 6 This is a schematic diagram illustrating the Jain's fairness index results for different methods provided in this embodiment. Detailed Implementation

[0039] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below. Technical features in various embodiments of the present invention can be combined accordingly without mutual conflict.

[0040] In the description of this invention, it should be understood that the terms "first" and "second" are used only for descriptive purposes and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined with "first" and "second" may explicitly or implicitly include at least one of those features.

[0041] like Figure 1 As shown, in a preferred embodiment of the present invention, the above-mentioned multi-UAV NOMA communication noise reduction method based on a pre-trained basic model includes the following steps S1 to S5. The specific implementation process of each step is described in detail below.

[0042] S1. A base model containing a multimodal embedding layer, a spatiotemporal graph location encoding layer, a Transformer encoder and a Transformer decoder, and a multi-task prediction head is pre-trained on a series of self-supervised or semi-supervised tasks. This enables the base model to learn the physical laws, spatiotemporal dynamic characteristics, and complex mapping relationships between the state and communication performance of the UAV NOMA network from UAV swarm operation data. After pre-training, the Transformer decoder in the base model is removed, and the result is used as the prediction model AeroNOMA.

[0043] It should be noted that, in order to enable the basic model to understand the physical laws of UAV NOMA networks, this invention S1 first constructs a pre-training dataset, which contains the operational data of UAV swarms across multiple consecutive time slots. Specifically, the pre-training dataset may contain a small amount of known, relatively optimal (but not globally optimal) UAV position information, power allocation coefficients, transmit power, and channel state information obtained through traditional optimization algorithms, constituting a semi-supervised learning signal. Secondly, the basic model is pre-trained based on MAE on tasks such as Masked Multi-Modal Autoencoding (MAE) and temporal prediction (including UAV motion state prediction and communication performance indicator trend prediction). Alternatively, the model can be trained or fine-tuned using an end-to-end reinforcement learning (RL) paradigm (such as PPO) for the policy network, or self-supervised pre-training can be performed using contrastive learning (SimCLR). Therefore, no limitation is made in this invention. In this embodiment, the masked multi-modal autoencoding task and the temporal prediction task are described below.

[0044] The Masked Multimodal Autoencoder (MAE) task borrows from the MAE concept. During the input phase, a portion of the input multimodal embedding tokens are randomly masked, and the encoder in the model only processes the visible tokens. Subsequently, a lightweight Transformer decoder receives the visible token features from the encoder output, along with shared masked tokens inserted at the masked locations. The Transformer decoder's task is to reconstruct the original masked token features. This task forces the model to learn the inherent correlations between data modalities.

[0045] The time series prediction task utilizes the Spatiotemporal Graph Position Encoding (STG-PE) mechanism in the model architecture and the Causal Masking mechanism in the decoder to enable the model to process time series data. It is primarily based on historical data. Based on the operational data of the first time slot, predict the future first time slot. The motion state of the UAV in each time slot (e.g., position vector) can also be established. Furthermore, the UAV NOMA network state (e.g., CSI, position, power) and communication performance (e.g., spectral efficiency SE, robustness) can also be established. A direct mapping between (NOMA network and other network representations). By pre-training the model, the performance metrics that the UAV NOMA network will achieve under specific states can be predicted, thereby guiding the model to learn "good" state representations.

[0046] Through the above tasks, the model learns the relationship between power allocation, channel state, and performance metrics, enabling it to implicitly learn how to perform power allocation and link selection to achieve joint noise cancellation (i.e., SINR optimization) for both uplink and downlink NOMA links. Furthermore, in this invention, the pre-training employs a hierarchical optimization strategy, separately optimizing the minimum signal-to-interference-plus-noise ratio (SINR) prediction, UAV location information, power allocation coefficients, transmit power, and channel state information. This can be replaced by a weighted summation of the aforementioned metrics to accommodate different preferences.

[0047] It should be noted that the basic model in this invention encompasses core components such as a multimodal embedding layer, a spatiotemporal graph location encoding layer, a Transformer encoder and decoder, and a multi-task prediction head. Each component has several alternative forms. In this embodiment, the multimodal embedding layer uses a parallel modality-specific MLP and a learnable context. The token can be replaced by using a cross-attention mechanism for early intermodal fusion, or by expanding the input modalities (such as battery level, QoS requirements); in this embodiment, the spatiotemporal graph location encoding layer uses spatiotemporal graph location encoding (STG-PE) to fuse node, spatial and temporal embeddings, which can be replaced by using a GNN layer to explicitly model the topology, or simplified to standard sine / cosine temporal encoding and relative location encoding; in this embodiment, the Transformer encoder uses a standard Transformer, which can be replaced by a linear Transformer (such as Performer) or a graph neural network (GNN); the Transformer decoder uses the MAE paradigm and is removed during inference, which can be replaced by retaining the decoder for autoregressive inference; the multi-task prediction head uses a global average pooling and parallel MLP prediction head, which can be replaced by directly using the output vector of the Context Token for prediction, or by using a hierarchical / cascaded prediction head structure.

[0048] S2. The heterogeneous data of the real-time acquisition of the UAV swarm is used as the input of the prediction model. The heterogeneous input data is mapped to a unified high-dimensional embedding space through a multimodal embedding layer. A learnable context token is introduced to aggregate global information, thereby forming the input sequence of the spatiotemporal graph location encoding layer.

[0049] It should be noted that in step S2 of the present invention, as Figure 2As shown, for each time slot, heterogeneous data of three modalities are acquired: Channel State Information (CSI), UAV Kinematics, and NOMA Grouping. The heterogeneous data of each modality is flattened into a one-dimensional vector and then passed through a multilayer perceptron corresponding to that modality to map the flattened vector to a high-dimensional embedding space, thereby obtaining the corresponding modality embedding token. A learnable parameterized token is then used as the context token, and the context token is concatenated with all modality embedding tokens to serve as the input sequence for the spatiotemporal graph location coding layer.

[0050] In this embodiment, the first Channel state information in each time slot Estimated channel gain including all links etc.; Kinematic state of unmanned aerial vehicles Location of all drones Speed, etc. NOMA grouping strategy Includes the current resource block allocation matrix Then, the heterogeneous data for each modality. It is flattened into a one-dimensional vector and passed through a multilayer perceptron (MLP) dedicated to this modality. Map it to a From the dimensional embedding space, we obtain the modal embedding token, denoted as... Specifically, in this embodiment, the mode embedding token corresponding to the channel state information is denoted as... The modal embedding token corresponding to the UAV's kinematic state is denoted as The modal embedding token corresponding to the NOMA grouping strategy is denoted as They are represented as follows:

[0051]

[0052]

[0053]

[0054] in, , , This represents the multi-layer perceptron corresponding to the channel state information, UAV kinematic state, and NOMA grouping strategy. This represents the operation of flattening multidimensional data into a one-dimensional vector.

[0055] Secondly, this embodiment introduces a learnable parameterized token, denoted as... This token is optimized using gradient descent during the pre-training phase to capture the global context of the task. This represents each modal embedding token after mapping (e.g.) , , Essentially, they are all of length 1. A real vector.

[0056] Finally, the context token is concatenated with all modality embedding tokens to form the input sequence of the spatiotemporal graph location encoding layer (before adding location encoding), denoted as . :

[0057]

[0058] in, This represents a vector concatenation operation; Represents the set of real numbers; Indicates the batch size during inference. This represents the total number of tokens in the input sequence. In this embodiment, ; This represents the feature dimension of the high-dimensional embedding space of the prediction model.

[0059] S3. In the spatiotemporal graph location encoding layer of the prediction model, the node identity encoding, spatial relationship encoding, and temporal evolution encoding of the UAV NOMA network state are concatenated and passed through a linear projection layer to obtain the spatiotemporal graph location encoding sequence, which is added to the input sequence as an embedding vector sequence that integrates spatiotemporal topological information.

[0060] It should be noted that in step S3 of the present invention, as Figure 2 As shown, to enable the prediction model to understand the dynamic physical topology of the UAV NOMA network, a spatiotemporal graph position encoding (STG-PE) mechanism is designed in the spatiotemporal graph position encoding layer. This mechanism comprehensively considers the identity information of UAV nodes, the relative spatial positional relationships between nodes, and the temporal evolution information of the UAV NOMA network state, thereby generating a spatiotemporal graph position encoded sequence with the same dimension as the input sequence. This is then injected into the input sequence. Specifically, the spatiotemporal map location encoding sequence consists of three parts:

[0061] Node identity encoding Assign a learnable embedding vector to each drone ID index among multiple drones, then for a token representing the relationship between two drone ID indices (such as...), (a component in the data), whose node identity is encoded by adding the embedding vectors corresponding to the two drone number indices, that is: for a drone number index and Its corresponding node identity code is , and Indexing drones by number and The embedding vector.

[0062] Spatial Relationship Coding : in the Within each time slot, the position difference between the two UAVs is normalized and used as a relative position vector. Then the relative position vector is input into a multilayer perceptron. Encoding is performed to obtain: .in, and They represent the first Driving drones and the The location of the drone.

[0063] Temporal Evolution Coding Index for each time slot Assign a learnable temporal embedding This time embedding is then assigned to all tokens in the current time slot input sequence, thus obtaining the time evolution code. .

[0064] Specifically, this embodiment first maintains a Temporal Embedding LookupTable, denoted as... ,in The maximum number of time slots is preset (e.g., for a UAV NOMA network with an operating cycle of 1000 time slots). ), The feature dimension is for time embedding. This lookup table is initialized using a Gaussian method during base model initialization and automatically optimized during training via gradient descent. For each time slot index... The corresponding time embedding is extracted from the time embedding lookup table through index operations. ,in, Indicates embedding lookup table from time No. row extraction A dimensional vector. Then, the extracted temporal embedding is... Assigning values ​​to all tokens in the current time slot yields the final time evolution code. This encoding is the same for all tokens within the same time slot and is used to identify their time location.

[0065] Furthermore, in this embodiment, the node identity encoding, spatial relationship encoding, and temporal evolution encoding of the UAV NOMA network state are concatenated and then projected through a linear projection layer composed of a multi-layer sensing mechanism. Map it to Dimensions, generating spatiotemporal graph location encoding sequences:

[0066]

[0067] Finally, the spatiotemporal graph location encoding sequence is added to the input sequence to obtain the embedding vector sequence. And used as input to the Transformer encoder:

[0068]

[0069] S4. After the embedded vector sequence is processed by the Transformer encoder, a global average pooling operation is performed on the feature sequence output by the Transformer encoder to aggregate the feature sequence into a high-dimensional global state vector, which is used to represent the global state of the UAV NOMA network in the current time slot.

[0070] It should be noted that in step S4 of this invention, the Transformer encoder adopts the standard Transformer architecture, and is composed of... It is made up of stacked identical Transformer blocks. This indicates the number of Transformer blocks in the Transformer encoder. The Transformer encoder captures the complex spatiotemporal dependencies between heterogeneous multimodal data and between different UAV nodes through a multi-head self-attention mechanism. After processing by several Transformer blocks, the output is a feature sequence containing rich contextual information. .

[0071] like Figure 2 As shown, for the first Transformer blocks ( ), its input is the first Output of a Transformer block The main computational processes involved in this Transformer block are as follows:

[0072] 1) Multi-Head Self-Attention (MHSA):

[0073]

[0074] in, Indicates the first The intermediate feature sequence after multi-head self-attention mechanism processing in a Transformer block; Presentation layer normalization operation; This represents a multi-head self-attention mechanism.

[0075] 2) Feed-Forward Network (FFN):

[0076]

[0077] in, Indicates the first The output of each Transformer block; This represents a feedforward neural network.

[0078] It should be noted that in step S4 of this invention, a global average pooling operation is used to... The characteristics of individual tokens are aggregated into a single... 3D global state vector This vector represents the global state of the entire drone swarm network in the current time slot, where, This indicates the Global Average Pooling operation.

[0079] S5. In the multi-task prediction head, the global state vector is fed into multiple prediction heads in parallel to predict the NOMA resource block grouping policy, the transmit power of each UAV, the three-dimensional position vector of each UAV, the minimum signal-to-interference-plus-noise ratio of all communication links in the UAV NOMA network under the current grouping policy, and the fairness index under the current grouping policy. Based on the predicted transmit power, the power allocation coefficient between UAVs is calculated, thereby achieving multi-objective joint optimization that maximizes spectrum efficiency, fairness and robustness.

[0080] It should be noted that in step S5 of this invention, the global state vector is fed into multiple dedicated prediction heads in parallel. Each prediction head consists of a multi-layer sensing mechanism and is used to predict a set of decision variables or performance indicators in a multi-objective optimization problem, including NOMA resource block grouping strategy, UAV transmit power, UAV position, minimum signal-to-interference-plus-noise ratio, and fairness indicators, thereby achieving joint optimization of communication resources such as the spectrum efficiency of UAV swarms.

[0081] In this embodiment, as Figure 2 As shown, the global state vector The data is fed in parallel into five independent prediction heads in the form of MLPs:

[0082] Grouping strategy prediction head take over It outputs the probability distribution representing the NOMA resource block grouping strategy. ,Right now A drone in The probability distribution across available resource blocks is used to determine the current resource block allocation matrix. Here, "1" represents the option of "do not activate resource blocks / do not transmit". This indicates the total number of drones.

[0083] Transmit power prediction head take over Output the transmit power of each drone. And constitutes a set of all the drones' transmit power. The predicted transmit power corresponds to the first... The drone in Actual transmit power of each time slot The actual transmission power is limited by , Indicates the lower bound of power. This indicates the upper limit of the maximum permissible transmit power.

[0084] Location prediction head take over Output the 3D position vector of each UAV, and form a set of the 3D position vectors of all UAVs. , Represents three-dimensional spatial coordinates. The predicted three-dimensional position vector corresponds to the... The drone in The actual location of each time slot It is used to guide drones in adjusting their position to optimize channel conditions.

[0085] Minimum signal-to-interference-plus-noise ratio prediction head take over Output the predicted minimum signal-to-interference-plus-noise ratio (SINR) for all communication links in the UAV NOMA network under the current grouping strategy. This output corresponds to the robustness metric of the UAV NOMA network. It is used to assess and ensure the communication quality of the weakest link in the drone NOMA network, ensuring that all drone nodes can maintain reliable communication.

[0086] Fairness prediction head take over Output a scalar value representing the fairness metric (JFI) under the predicted grouping strategy. This scalar value can be used to assist training or as a real-time monitoring indicator of the performance of the UAV NOMA network.

[0087] Finally, the result set output by the above prediction head and the power allocation coefficients implicitly included in the prediction model This is the solution sought in this invention to a multi-objective optimization problem that aims to jointly maximize spectral efficiency, fairness, and robustness, thereby achieving real-time communication resource optimization for UAV swarms.

[0088] The above power allocation coefficient In a NOMA network for unmanned aerial vehicles (UAVs), when multiple users share the same time-frequency resource block, in the case of the [missing information], [missing information] On the resource block, the first Driving drones assigned to the first The power ratio of the UAV communication link. This coefficient is not a direct output of the prediction model, but can be indirectly derived from the explicit output of the prediction model. Specifically, it is derived from the probability distribution of the predicted NOMA resource block grouping strategy. It can determine which drones share the same resource block and predict the set of transmit powers. Each element in the equation represents the transmit power of each UAV, and channel state information (input) determines the channel gain order of the link. Based on the NOMA serial interference cancellation (SIC) principle, the power allocation coefficients can be derived according to the following rules:

[0089]

[0090] in, In NOMA systems, at the receiver, serial interference cancellation (SIC) technology and commonly used fixed power allocation methods are used to decode each user's signal sequentially. Typically, users with poorer channel conditions are allocated higher power to ensure that their signals can be successfully decoded. For channel gain, the SIC order is usually arranged from smallest to largest. Users with poorer channel conditions receive larger power allocation coefficients. Therefore, based on the power allocation coefficients mentioned above, if the... The transmit power of the drone is Then it is assigned to the link. The actual power is:

[0091]

[0092] The following constraints must be satisfied:

[0093]

[0094]

[0095] To better demonstrate the specific implementation and technical effects of the present invention, the multi-UAV NOMA communication noise reduction method based on the pre-trained basic model shown in steps S1 to S5 of the above preferred implementation is applied to a specific example.

[0096] Example

[0097] The specific implementation process of the multi-UAV NOMA communication noise reduction method based on the pre-trained basic model used in this embodiment is as described above and will not be repeated here. This invention achieves maximum spectral efficiency (SE), fairness, and robustness through joint optimization of UAV position, transmit power, power allocation coefficient, and NOMA resource block grouping.

[0098] This invention demonstrates excellent scalability and robustness in high-density, high-complexity network environments: such as Figure 3 As shown, compared to baseline algorithms such as Particle Swarm Optimization (PSO), Proximity Policy Optimization (PPO), and Graph Neural Networks (GNN), the robustness metric (minimum SINR) of baseline algorithms rapidly decreases with the increase in the number of drones. However, this invention, benefiting from effective joint management of uplink, downlink, and self-interference, maintains a stable minimum SINR even when the number of drones increases from 7 to 13, while the baseline algorithms deteriorate drastically. Figure 4 As shown, the total spectral efficiency (Total SE) of this invention continues to increase even with 13 drones, without reaching a performance inflection point; while the baseline algorithm's performance saturates or begins to decline with 8 to 10 drones, demonstrating the robust optimization capability of this invention under high load. For example, in a scenario with 13 drones, the total spectral efficiency of this invention can reach over 42 bps / Hz, while the best-performing GNN baseline algorithm is below 35 bps / Hz, representing an improvement of approximately 14%. Figure 5As shown, compared with baseline algorithms such as Particle Swarm Optimization (PSO), Proximity Policy Optimization (PPO), and Graph Neural Networks (GNN), the average spectral efficiency (Average SE) of this invention is significantly improved. Figure 6 As shown, the Jain's fairness index of this invention consistently remains at a high level above 0.93s, significantly outperforming all baseline algorithms whose performance degrades with the increase in the number of drones. In summary... Figures 3-6 The results show that the method of the present invention has significant advantages over other baseline algorithms.

[0099] It is understood that the multi-UAV NOMA communication denoising method based on a pre-trained basic model described in S1-S5 above can essentially be implemented by a computer program. Therefore, based on the same inventive concept, another preferred embodiment of the present invention also provides a computer program product corresponding to the multi-UAV NOMA communication denoising method based on a pre-trained basic model provided in the above embodiments. This product includes a computer program / instruction, which, when executed by a processor, can implement the multi-UAV NOMA communication denoising method based on a pre-trained basic model as described in the above embodiments.

[0100] Similarly, based on the same inventive concept, another preferred embodiment of the present invention also provides a computer electronic device corresponding to the multi-UAV NOMA communication noise reduction method based on the pre-trained basic model provided in the above embodiment, which includes a memory and a processor;

[0101] The memory is used to store computer programs;

[0102] The processor is configured to implement the multi-UAV NOMA communication noise reduction method based on a pre-trained basic model in the above embodiments when executing the computer program.

[0103] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0104] Therefore, based on the same inventive concept, another preferred embodiment of the present invention also provides a computer-readable storage medium corresponding to the multi-UAV NOMA communication noise reduction method based on a pre-trained basic model provided in the above embodiments. The storage medium stores a computer program, which, when executed by a processor, can realize the multi-UAV NOMA communication noise reduction method based on a pre-trained basic model in the above embodiments.

[0105] Specifically, in the computer-readable storage medium of the above three embodiments, the stored computer program is executed by a processor, which can perform the aforementioned steps S1 to S5.

[0106] It is understood that the aforementioned storage media may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Furthermore, the storage media may also be various media capable of storing program code, such as USB flash drives, external hard drives, magnetic disks, or optical discs.

[0107] It is understood that the processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0108] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the invention. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the invention. Therefore, all technical solutions obtained through equivalent substitution or transformation fall within the protection scope of the present invention.

Claims

1. A method for denoising multi-UAV NOMA communication based on a pre-trained base model, characterized in that, Includes the following steps: S1. A base model containing a multimodal embedding layer, a spatiotemporal graph location encoding layer, a Transformer encoder and a Transformer decoder, and a multi-task prediction head is pre-trained on a series of self-supervised or semi-supervised tasks. This enables the base model to learn the physical laws, spatiotemporal dynamic characteristics, and complex mapping relationship between the state and communication performance of the UAV NOMA network from UAV swarm operation data. After pre-training, the Transformer decoder in the base model is removed and used as the prediction model. S2. The heterogeneous data of the real-time acquisition of the drone swarm is used as the input of the prediction model. The heterogeneous data input is mapped to a unified high-dimensional embedding space through a multimodal embedding layer. A learnable context token is introduced to aggregate global information, thereby forming the input sequence of the spatiotemporal graph location encoding layer. S3. In the spatiotemporal graph location encoding layer of the prediction model, the node identity encoding, spatial relationship encoding and time evolution encoding of UAV NOMA network state of the UAV are concatenated and passed through a linear projection layer to obtain the spatiotemporal graph location encoding sequence, which is added to the input sequence as the embedding vector sequence that integrates spatiotemporal topology information. S4. After the embedded vector sequence is processed by the Transformer encoder, a global average pooling operation is performed on the feature sequence output by the Transformer encoder to aggregate the feature sequence into a high-dimensional global state vector, which is used to represent the global state of the UAV NOMA network in the current time slot. S5. In the multi-task prediction head, the global state vector is fed into multiple prediction heads in parallel to predict the NOMA resource block grouping policy, the transmit power of each UAV, the three-dimensional position vector of each UAV, the minimum signal-to-interference-plus-noise ratio of all communication links in the UAV NOMA network under the current grouping policy, and the fairness index under the current grouping policy. Based on the predicted transmit power, the power allocation coefficient between UAVs is calculated, thereby achieving multi-objective joint optimization that maximizes spectrum efficiency, fairness and robustness.

2. The multi-UAV NOMA communication noise reduction method based on a pre-trained base model as described in claim 1, characterized in that, In step S2, for each time slot, heterogeneous data of three modes are acquired, namely channel state information, UAV kinematic state information, and NOMA grouping strategy. The heterogeneous data of each mode is flattened into a one-dimensional vector and then passed through a multilayer perceptron corresponding to the mode to map the flattened vector to a high-dimensional embedding space to obtain the corresponding mode embedding token. Then, a learnable parameterized token is used as the context token, and the context token is concatenated with all mode embedding tokens as the input sequence of the spatiotemporal graph location coding layer.

3. The multi-UAV NOMA communication noise reduction method based on a pre-trained base model as described in claim 1, characterized in that, In step S3, a learnable embedding vector is assigned to each drone number index among the multiple drones. Then, for a Token representing the relationship between two drone number indices, its node identity encoding is the sum of the embedding vectors of the corresponding two drone number indices.

4. The multi-UAV NOMA communication noise reduction method based on a pre-trained base model as described in claim 1, characterized in that, In step S3, at the first In each time slot, the position difference between the two UAVs is normalized and used as a relative position vector. The relative position vector is then input into a multilayer perceptron for encoding to obtain the spatial relationship code.

5. The multi-UAV NOMA communication noise reduction method based on a pre-trained base model as described in claim 1, characterized in that, In step S3, index each time slot. A learnable temporal embedding is assigned and then assigned to all tokens in the current time slot input sequence to obtain the temporal evolution code.

6. The multi-UAV NOMA communication noise reduction method based on a pre-trained base model as described in claim 1, characterized in that, In step S5, the grouping policy prediction head receives the global state vector and outputs the probability distribution representing the NOMA resource block grouping policy; the transmit power prediction head receives the global state vector and outputs the transmit power of each UAV, forming a first set of transmit powers for all UAVs; the position prediction head receives the global state vector and outputs the three-dimensional position vector of each UAV, forming a second set of three-dimensional position vectors for all UAVs; the minimum signal-to-interference-plus-noise ratio (SINNR) prediction head receives the global state vector and outputs the predicted minimum SINNR value for all communication links in the UAV NOMA network under the current grouping policy; and the fairness prediction head receives the global state vector and outputs a scalar value representing the fairness index predicted under the current grouping policy.

7. The multi-UAV NOMA communication noise reduction method based on a pre-trained base model as described in claim 6, characterized in that, The grouping strategy prediction head, transmit power prediction head, position prediction head, minimum signal-to-interference-plus-noise ratio prediction head, and fairness prediction head are all based on a multilayer perceptron.

8. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it can implement the multi-UAV NOMA communication noise reduction method based on a pre-trained basic model as described in any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the multi-UAV NOMA communication noise reduction method based on a pre-trained basic model as described in any one of claims 1 to 7.

10. A computer electronic device, characterized in that, Including memory and processor; The memory is used to store computer programs; The processor is configured to, when executing the computer program, implement the multi-UAV NOMA communication noise reduction method based on a pre-trained basic model as described in any one of claims 1 to 7.

Images