A multi-UAV joint coverage optimization method based on Set Transformer design

Abstract

A multi-UAV joint coverage optimization method based on Set Transformer design relates to wireless communication network and UAV communication. A multi-ABS coverage optimization model based on Set Transformer is proposed, the image encoder, prompt encoder and mask decoder in the SAM model are reconstructed, the position encoder adopts the Set Transformer module to encode the coordinate information, the prompt encoder introduces the PLE module to process the scalar parameters, and the mask decoder fuses features through variable output tokens and attention mechanism and outputs the optimal position of ABS; the attention mechanism is used to encode the variable number of ABS and GU features, and the interaction relationship and spatial distribution characteristics of ABS and GU are captured. The forward-backward network training process is designed to enhance the adaptive ability of the model to different ABS numbers, coverage radii and GU spatial distributions, and to realize intelligent optimization selection of multi-ABS deployment position.

Description

Technical Field

[0001] This invention relates to the fields of wireless communication networks and UAV communication technology, and in particular to a multi-UAV joint coverage optimization method based on SetTransformer design. Background Technology

[0002] In numerous emergency situations, unmanned aerial vehicles (UAVs) can act as aerial base stations (ABS) to provide wireless communication services to ground users (GUs). However, achieving maximum coverage for M GUs by deploying N ABSs with limited coverage is an extremely challenging multi-UAV joint coverage problem. Essentially, this task is an NP-hard problem, with its complexity exponentially related to the values ​​of M and N.

[0003] Currently, existing literature on multi-UAV joint coverage problems mostly uses traditional heuristic algorithms. The Transformer (Vaswani A, Shazeer N, Parmar N, et al. Attention is all you need[J]. Advances in neural information processing systems, 2017, 30), since its introduction in 2017, has become a core architecture in the field of artificial intelligence, widely applied in natural language processing, computer vision, and other fields. Its design based on a self-attention mechanism addresses the limitations of traditional recurrent neural networks in long-range dependencies and parallel computation. The Transformer's advantages lie in its parallel computing capabilities and effective modeling of long-range dependencies. By adopting a self-attention mechanism, all input data can be processed in parallel, significantly improving the model's training efficiency. Therefore, using the Transformer to solve multi-UAV joint coverage problems has significant practical implications. Meanwhile, the SAM model is a fundamental model for image segmentation, designed to perform interactive segmentation with prompts and generalize to new image distributions and tasks in zero-shot cases. This model can be divided into three key parts: an image encoder, a prompt encoder, and a lightweight mask decoder. The core of the SAM model is the Transformer architecture, particularly its self-attention and cross-attention mechanisms in the image encoder and mask decoder. These mechanisms combine image embeddings and cue embeddings to generate high-quality segmentation masks. Due to its good generalization performance, the SAM model can be widely applied to various downstream problems (such as multi-UAV joint coverage optimization). Therefore, constructing an ST-SAM model based on Set Transformer to solve the multi-UAV joint coverage optimization problem is of great significance. Summary of the Invention

[0004] The purpose of this invention is to provide a multi-UAV joint coverage optimization method based on SetTransformer design for the multi-UAV joint coverage optimization problem.

[0005] This invention includes the following steps:

[0006] 1) System Model Construction and Problem Analysis: Define a wireless communication system containing multiple airborne base stations (ABS) and multiple ground users (GU), where the location set of GU is a known two-dimensional coordinate; based on the line-of-sight (LoS) channel model and free space path loss model, set the maximum ground coverage radius of the ABS at a given altitude, corresponding to the signal-to-noise ratio (SNR) threshold, and define the coverage index of the GU; model the optimization problem.

[0007] 2) Design the ST-SAM model. By introducing the Set Transformer module to build a position encoder, the Piecewise Linear Embedding (PLE) module to build a cue encoder, and introducing a variable number of output tokens and combining them with an internal attention mechanism to build a mask decoder, the SAM model is reconstructed to achieve encoding and decoding of relevant information and output of the optimal position of ABS.

[0008] 3) Forward network training: Input randomly generated GU coordinates, ABS coordinates, ABS coverage radius, and number of ABS; input the GU coordinates into the position encoder to generate a position embedding, and input the ABS coverage radius, number, and ABS position information processed by the position encoder into the cue encoder to generate a cue embedding; input the position embedding and cue embedding into the mask decoder, and output the predicted GU coverage marker (0 / 1 sequence); use the true coverage marker under the current ABS placement conditions as the label, and update the network parameters through mean squared error loss (MSELoss) and backpropagation;

[0009] 4) Reverse network training: Input GU coordinates, ABS coverage radius and number; input GU coordinates into the position encoder to generate position embedding, input ABS coverage radius and number into the cue encoder to generate cue embedding; initialize the preset maximum number of output tokens in the mask decoder, concatenate with the cue embedding and input into the decoder; perform feature interaction and fusion through the cross attention and multi-head self attention mechanism of the Transformer module, and output the predicted optimal ABS position coordinates according to the target ABS number;

[0010] 5) Joint training of forward and backward networks: After the forward network is fully trained, all its parameters are frozen as a fixed coverage calculation module; when training the backward network, the ABS position coordinates output by the backward network are used to generate position codes through the frozen position encoder, and input into the prompt encoder of the forward network to calculate the corresponding GU coverage; with "all GUs are covered" as the ideal label, the error between the prediction result of the forward network and the ideal label is used to backpropagate and optimize the parameters of the backward network until the backward network stably outputs the optimal ABS position;

[0011] 6) Output optimization results: Based on the trained ST-SAM model, input the target GU location, ABS number and coverage radius, and quickly output the optimal ABS prediction location to maximize GU coverage.

[0012] In step 1), the specific steps for constructing the system model and analyzing the problem can be as follows: Consider a wireless communication system with N ABS and M GU, where The location set of GU {q m} m∈M Given that q m ∈R 2×1 Let represent the two-dimensional coordinates of the m-th GU on the ground; assume that the ABS-GU communication channel is mainly dominated by the LoS link; also assume that the transmit power is fixed and the minimum signal-to-noise ratio (SNR) required for reliable communication at the receiver is known; under the LoS model, the power gain of the ABS-GU channel follows the free-space path loss model, which is determined by the ABS-GU distance; to explore an algorithm that approximates the optimum in a finite time, assume that the ABS flies at a given altitude H, and its maximum coverage radius R on the ground corresponds to the SNR threshold; define the coverage index of GU m as:

[0013]

[0014] Where, p n ∈R 2×1 Represents the horizontal coordinate of ABSn, ||p n -q m || represents the distance between the projections of GUm and ABSn onto the ground;

[0015] The goal of the multi-UAV joint coverage optimization problem is to maximize the coverage of the GU using N ABS (Automatic Protective Units). Therefore, the problem is modeled as (P1):

[0016]

[0017] Among them, C m q is the coverage metric for GU m. m ∈R 2×1 p represents the two-dimensional coordinates of the m-th GU on the ground. n ∈R 2×1 Represents the horizontal coordinate of ABSn, ||p n -q m || represents the distance between the projections of GUm and ABSn onto the ground, and R is the maximum coverage radius on the ground.

[0018] In step 2), the ST-SAM model is designed by first improving the original image encoder by introducing a SetTransformer to replace its original module with a position encoder. This position encoder encodes the GU map information to improve the model's ability to extract user location features and significantly reduce computational resource consumption. The cue encoder is redesigned by introducing a PLE module to enable the model to encode input scalar parameters, such as ABS coverage radius, more reasonably and efficiently. A variable number of output tokens are introduced into the mask decoder, and the attention mechanism inside the mask decoder is used to achieve the optimal position output for the variable number of ABS.

[0019] In step 3), the specific steps of training the forward network are as follows: In the forward network, the input data includes randomly generated GU coordinates, ABS coordinates, ABS coverage radius R, and the number of ABS. The training labels are user coverage flags under the current ABS placement conditions. The user coverage flag is a sequence consisting only of 0 / 1. First, the GU location coordinate information is input into the model's location encoder, which can effectively encode all GU locations on the map to form corresponding location embeddings, which serve as one of the inputs to the subsequent mask decoder. Then, the ABS coverage radius and the number of ABS are used as prompts and encoded into the prompt encoder to obtain the prompt embedding. The ABS location information is also processed by the location encoder to obtain the ABS location embeddings, which are then input into the prompt encoder. Finally, the location embeddings and prompt embeddings are input into the improved mask decoder. The cross-attention and self-attention modules in the decoder output the predicted user coverage flags. The corresponding labels are the pre-calculated actual user coverage obtained based on the current ABS placement. Then, the mean squared error loss (MSE) between the two is calculated. The network performs a backpropagation (BackPropagation) to update the network parameters.

[0020] In step 4), the specific steps of training the reverse network are as follows: In the reverse network, firstly, the GU coordinate information is input into the position encoder to obtain the corresponding GU position embedding; simultaneously, the coverage radius and number of ABS are used as cues to input into the cue encoder to obtain the corresponding cue embedding; next, when initializing the mask decoder, a special maximum number (the maximum number of ABS is set to 5 in this invention) of output tokens is pre-embedded for subsequent output of a variable number of optimal UAV positions; when the mask decoder receives the GU position embedding and cue embedding, the output token is concatenated with the cue embedding to obtain the concatenated output token, which is then input into the improved mask decoder; the Transformer module in the decoder uses cross-attention and multi-head attention mechanisms to perform feature interaction and feature fusion between the concatenated output token and the GU position embedding; this mechanism enables the model to simultaneously focus on the GU position information and the input cue information, thereby ultimately accurately predicting and outputting the optimal ABS position coordinates according to the required number of ABS output.

[0021] In step 5), the specific steps of the joint training of the forward and reverse networks are as follows: After sufficient training in the early stage, the forward network can accurately fit an efficient and reliable calculation function, that is, calculate the GU coverage in real time according to the ABS position; when actually training the reverse network, firstly, all parameters of the pre-trained forward network are frozen as a fixed coverage calculation module; during the training process, the specific process of the reverse network is as follows: firstly, the GU coordinate information is input into the position encoder to generate the GU position embedding, and at the same time, the ABS coverage radius and ABS quantity information are input into the cue encoder to generate the cue embedding; then, the reverse network outputs the optimal ABS position coordinate prediction result through the improved mask decoder, and these position coordinates generate position encoding information through the frozen position encoder; this position encoding information is then input into the cue encoder in the forward network as part of the cue information to assist the forward network in accurately calculating the GU coverage under the corresponding ABS position; throughout the training process, the ideal label for GU coverage is set to all users being covered (i.e., the corresponding label maxCoverage). The Flags sequence value is set to 1 to maximize ground user coverage. By calculating the error between the coverage predicted by the forward network and the ideal label, the parameters of the reverse network are optimized by back propagation, and the optimization is continuously iterated. After the above iterative training, the reverse network can finally accurately output the optimal ABS deployment location, thereby maximizing the optimization of ground GU coverage.

[0022] This invention constructs an ST-SAM model based on Set Transformer. First, the original image encoder is improved by introducing a Set Transformer to replace its original module with a position encoder. Second, the cue encoder is redesigned, introducing a PLE module to enable the model to encode input scalar parameters, such as ABS coverage radius, more reasonably and efficiently. Furthermore, a variable number of output tokens are introduced into the mask decoder, and the optimal position output for the variable number of ABS is achieved using the mask decoder's internal attention mechanism. Simultaneously, based on the SAM model architecture, this invention designs a forward-backward training network.

[0023] Compared with the prior art, the present invention has the following outstanding advantages:

[0024] 1. This invention uses the Set Transformer module to replace the encoder in the SAM model. This location encoder can more effectively encode the location of GU map information, thereby improving the model's ability to extract GU location features and significantly reducing computational resource consumption.

[0025] 2. This invention introduces a PLE module into the prompt encoder, enabling the model to encode the input scalar parameters more reasonably and efficiently.

[0026] 3. This invention introduces a variable number of output tokens into the mask decoder and utilizes the attention mechanism inside the mask decoder to achieve the optimal position output of the variable number of ABS.

[0027] 4. To further enhance the model's ability to understand input information and improve its performance in selecting ABS locations, this invention designs a forward-backward training network. This network not only effectively assists the model training process but also improves the efficiency of computing resource utilization and the convenience of training.

[0028] 5. Regarding GU coverage performance, this invention approaches the theoretical performance upper bound obtained using the grid exhaustive search method (GES) in small-scale scenarios, while significantly reducing computation time. Furthermore, the GU coverage performance of this invention outperforms other benchmark solutions. Attached Figure Description

[0029] Figure 1 This is a diagram of the ST-SAM model architecture of the present invention.

[0030] Figure 2 This is a flowchart of the forward network of the present invention.

[0031] Figure 3 This is a reverse network flowchart of the present invention.

[0032] Figure 4 This is a flowchart of the forward-backward network training process of the present invention.

[0033] Figure 5 This is a flowchart of the position encoder of the present invention.

[0034] Figure 6 The encoder flowchart is shown for this invention.

[0035] Figure 7 This is a flowchart of the forward network mask decoder of the present invention.

[0036] Figure 8 This is a flowchart of the reverse network mask decoder of the present invention.

[0037] Figure 9 Optimal location prediction for ABS with variable numbers. Where a to d correspond to 1 to 4 ABS respectively.

[0038] Figure 10 ABS optimal location prediction for various algorithms. Among them, a is the result of the GES exhaustive search method, b is the result of the proposed ST-SAM method, c is the result of TD3, d is the result of Kmeans, e is the result of DBSCAN, and f is the result of the random method.

[0039] Figure 11 Coverage analysis of various algorithms under variable ABS and GU numbers. The subgraphs correspond to cases with GU numbers ranging from 100 to 500, from top to bottom.

[0040] Figure 12 Coverage analysis of various algorithms under variable ABS quantity and ABS coverage radius. Among them, each sub-map corresponds to a coverage radius of 50-100 meters, 100-150 meters, and 150-200 meters from top to bottom. Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of this invention clearer, the following embodiments will be used in conjunction with the accompanying drawings to further illustrate the invention. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0042] This invention proposes an ST-SAM model based on Set Transformer (Lee J, Lee Y, Kim J, et al. Settransformer: A framework for attention-based permutation-invariant neural networks [C] / / International conference on machine learning.PMLR,2019:3744-3753) to solve the multi-UAV joint coverage optimization problem. The ST-SAM model architecture diagram of this invention is shown below. Figure 1 As shown. This invention reconstructs the image encoder, cue encoder, and mask decoder in the SAM (Kirillov A, Mintun E, Ravi N, et al. Segment anything[C] / / Proceedings of the IEEE / CVF international conference on computer vision.2023:4015-4026) model. A Set Transformer module is introduced to construct the position encoder, and a PLE (Gorishniy Y, Rubachev I, Babenko A. On embeddings for numerical features in tabular deep learning[J]. Advances in Neural Information Processing Systems,2022,35:24991-25004) module is introduced to construct the cue encoder. Simultaneously, a variable number of output tokens are introduced into the mask decoder, enabling the model to encode and decode a variable number of ABS, GU, and a variable coverage radius, quickly outputting the optimal ABS prediction position.

[0043] The embodiments of the present invention include the following steps:

[0044] 1) Constructing a system model and analyzing problems

[0045] This invention considers a wireless communication system with N ABS and M GU, wherein The location set of GU {q m ) m∈M Given that q m ∈R 2×1Let represent the two-dimensional coordinates of the m-th GU on the ground. To simplify the problem, it is assumed that the ABS-GU communication channel is mainly dominated by a LoS link. It is also assumed that the transmit power is fixed and the minimum signal-to-noise ratio (SNR) required for reliable communication at the receiver is known. Under the LoS model, the power gain of the ABS-GU channel follows a free-space path loss model, which is determined by the ABS-GU distance. Since the problem studied in this invention is NP-hard, to explore an algorithm that approximates the optimum within a finite time, it is assumed that the ABS flies at a given altitude H, and its maximum coverage radius R on the ground corresponds to the SNR threshold. Therefore, the coverage index of GU m is defined as:

[0046]

[0047] Where p n ∈R 2×1 Represents the horizontal coordinate of ABSn, ||p n -q m || represents the distance between the projections of GU m and ABSn onto the ground.

[0048] The problem studied in this invention is the joint coverage optimization problem of multiple UAVs, and the goal is to maximize the coverage of the GUs using N ABSs. Therefore, this invention models the problem as (P1):

[0049]

[0050] 2) Forward-Reverse Network Architecture

[0051] The model structure process proposed in this invention is as follows: Figure 1 As shown in the figure, the SAM model underwent structural adjustments and optimizations, resulting in the ST-SAM model. First, the original image encoder was improved by introducing a Set Transformer to replace its original module with a position encoder. This position encoder can more effectively encode the location of GU map information, thereby improving the model's ability to extract user location features while significantly reducing computational resource consumption. Second, the cue encoder was redesigned by introducing a PLE module, enabling the model to encode input scalar parameters, such as ABS coverage radius, more reasonably and efficiently. Furthermore, this invention introduces a variable number of output tokens into the mask decoder and utilizes the mask decoder's internal attention mechanism to achieve optimal position output for the variable number of ABS.

[0052] 2.1) Forward Network Training Process

[0053] In step 2.1), the specific method can be as follows: Based on the given ABS coordinates, the coverage of the GU is calculated in real time. Specifically, a neural network is used to fit the coverage solution process, and its specific flow structure is as follows: Figure 2 As shown. The network input data includes randomly generated GU coordinates, ABS coordinates, ABS coverage radius R, and the number of ABS. The training labels are user coverage flags under the current ABS placement conditions. The user coverage flag is a sequence consisting only of 0 / 1. The specific training process of the forward network is as follows: First, in the forward network, the GU position coordinate information is input to the model's position encoder. This position encoder can effectively encode all GU positions on the map, forming corresponding position embeddings, which serve as one of the inputs to the subsequent mask decoder. Subsequently, the ABS coverage radius and the number of ABS are used as prompt inputs to the prompt encoder for encoding, resulting in a prompt embedding. The ABS position information is also processed by the position encoder to obtain the ABS position embeddings, which are then input into the prompt encoder. Finally, the aforementioned location embedding and cue embedding are input into the improved mask decoder. The cross-attention and self-attention modules in the decoder output the predicted user coverage flags, and the corresponding labels are the real user coverage obtained in advance based on the current ABS placement. Then, the mean squared error loss (MSELoss) between the two is calculated, and back propagation is performed to update the network parameters.

[0054] 2.2) Reverse Network Training Process

[0055] In step 2.2), the specific process structure of the reverse network is as follows: Figure 3As shown, the training process is described as follows: In the inverse network, the GU coordinate information is first input into the position encoder to obtain the corresponding GU position embedding. Simultaneously, information such as the coverage radius and number of ABS are input as cues to the cue encoder to obtain the corresponding cue embedding. Next, during the initialization of the mask decoder, a special maximum number (in this invention, the maximum number of ABS is set to 5) of output tokens is pre-embedded for subsequent output of a variable number of optimal UAV positions. When the mask decoder receives the GU position embedding and cue embedding, it concatenates the output token with the cue embedding to obtain the concatenated output token, which is then fed into the improved mask decoder. The Transformer module within the decoder uses cross-attention and multi-head attention mechanisms to perform feature interaction and feature fusion between the concatenated output token and the GU position embedding. This mechanism enables the model to simultaneously focus on the GU position information and the input cue information, thereby accurately predicting and outputting the optimal ABS position coordinates according to the required number of ABS outputs.

[0056] 2.3) Joint Training Process of Forward and Backward Networks

[0057] In step 2.3), Figure 4This describes the joint training process of the forward and backward networks. The dashed lines represent the parameter modules that are frozen, while the solid lines represent the modules that update the parameters. Specifically, after sufficient pre-training, the forward network can accurately fit an efficient and reliable calculation function, namely, calculating GU coverage in real time based on ABS positions. When training the backward network, all parameters of the pre-trained forward network are first frozen as fixed coverage calculation modules. The specific process of the backward network during training is as follows: First, the GU coordinate information is input into the position encoder to generate GU position embeddings. Simultaneously, the ABS coverage radius and ABS quantity information are input into the cue encoder to generate cue embeddings. Then, the backward network outputs the optimal ABS position coordinate prediction results through an improved mask decoder. These position coordinates are used by the frozen position encoder to generate position encoded information. This position encoded information is then input into the cue encoder in the forward network as part of the cue information, assisting the forward network in accurately calculating the GU coverage at the corresponding ABS position. Throughout the training process, the ideal label for GU coverage is set to ensure all users are covered (i.e., the corresponding maxCoverage Flags sequence value is set to 1), aiming to maximize ground user coverage. By calculating the error between the coverage predicted by the forward network and the ideal label, the parameters of the reverse network are optimized through backpropagation, iteratively improving the network. After repeated iterative training, the reverse network can ultimately accurately output the optimal ABS deployment location, thereby maximizing ground GU coverage.

[0058] 3) Position encoder

[0059] Step 3) will detail the module structure and specific functions of the position encoder, as shown in the flowchart below. Figure 5 As shown. The input dimension of the position encoder module is K*2, and the output dimension is K*D. The input user or drone coordinate information is first processed by the position encoding module. Furthermore, since position encoding is based on the literature (Vaswani A, Shazeer N, Parmar N, et al. Attention is all you need[J]. Advances in neural information processing systems, 2017, 30), it is inherently not limited by sequence length. Therefore, the model still exhibits excellent generalization performance even when processing sequences longer during the inference phase than during the training phase. The specific formula for position encoding is:

[0060]

[0061] Where pos is the index position of the element in the sequence, i is the dimension index, and d is the position of the element. model It is the embedding dimension, and this encoding method allows the model to learn relative position information.

[0062] After the positional encoding module, the positional encoder undergoes two consecutive ISAB (InducedSetAttention Block) modules. ISAB is a self-attention mechanism module specifically designed for set data, aiming to effectively reduce the computational complexity of traditional self-attention mechanisms. Specifically, the embedded information obtained after positional encoding can be considered as set data, and the ISAB module is naturally adapted to such set inputs, performing particularly well in scenarios where the number of input set elements is variable. ISAB introduces a set of inducing points as "information hubs" for the entire set's information interaction. This significantly reduces the computational complexity of the self-attention mechanism while ensuring that its processing is unaffected by changes in the order or number of input elements. Specifically, these inducing points effectively aggregate the global information of the entire set and propagate this global information to each input element through the attention mechanism. This allows each element of the input set to focus on local information while effectively acquiring the global features of the entire set. Furthermore, ISAB possesses permutation equivariance, meaning that its output adjusts accordingly regardless of the arrangement of the input set elements, enabling the model to process coordinate set data more flexibly and robustly. Through the introduction of the ISAB module, the model can not only handle larger-scale input sets more effectively but also enhance its global modeling ability of the input data. This mechanism of ISAB also effectively reduces the position encoder's over-reliance on local input information, preventing the risk of overfitting. Finally, through the continuous processing of the position encoding and the ISAB module, the coordinate information is successfully mapped from the initial K*2 dimension to a higher-dimensional K*D space. In this process, the ISAB module efficiently captures the detailed features and higher-order relationships between coordinate position data through a self-attention mechanism, thereby improving the model's overall understanding of spatial position features and its generalization performance.

[0063] 4) Prompt encoder

[0064] Step 4) will detail the module structure and specific function of the prompt encoder, as shown in the structural diagram below. Figure 6As shown. This invention structurally modifies the cue encoder in the SAM model by introducing a Piecewise Linear Embeddings (PLE) module. The PLE module can more effectively encode continuous or discrete scalar values ​​with finer detail, enabling the model to fully understand the meaning and characteristics of these scalar parameters, thereby better guiding the model's attention mechanism. The cue encoder module mainly consists of three sub-modules, as follows: Figure 6 As shown, the core module is the Piecewise Linear Embeddings (PLE) module. Specifically, PLE first divides the value range of each numerical feature into several non-overlapping bins, each bin representing a local pattern of the feature. When a feature value falls into a specific bin, PLE generates a continuous embedding vector through linear interpolation to capture the local changes of the numerical feature in different bins. Specifically, in this invention, for scalar information such as the ABS coverage radius and the number of ABS, the PLE module can accurately capture the inherent feature patterns contained in these scalar values. In addition, the PLE module incorporates learnable parameters, which are continuously updated through backpropagation to further optimize the embedding vector and achieve the best feature representation effect. The high-dimensional vector generated after encoding by the PLE module not only enhances the expressive power of the features but also effectively preserves the order of numerical features, while improving the robustness of the model to input changes. Subsequently, these vectors are further processed through linear layers and activation functions to form the final cue embedding vector. The resulting cue tokens are better suited for multi-ABS coverage optimization problems, effectively guiding the model to achieve efficient attention focusing and feature fusion.

[0065] 5) Mask Decoder

[0066] The flowchart of the mask decoder is as follows Figure 7 and Figure 8 As shown. Among them, Figure 7 This diagram illustrates the workflow of a mask decoder in a forward network. Figure 8 This diagram illustrates the workflow of the mask decoder in the inverse network. The mask decoder in the forward network combines the GU's positional encoding, cue encoding (ABS positional encoding, ABS count, and ABS coverage radius), passes it through the decoder's attention mechanism, and outputs the model's predicted GU coverage. The mask decoder in the inverse network combines the GU's positional encoding, cue encoding (ABS count and ABS coverage radius), and output token, passes it through the decoder's attention mechanism, and outputs the predicted ABS position.

[0067] 5.1) Workflow of the Mask Decoder in the Forward Network. The input to the mask decoder in the forward network mainly consists of two parts: position embeddings and cue tokens. The specific workflow can be divided into the following stages: The first stage is the self-attention stage of the cue tokens. The mask decoder first performs a self-attention operation on the input cue tokens. This means that each element in the cue information will interact with other cue elements, thereby capturing the correlation between cue tokens. The second stage is the cross-attention stage of the cue tokens and position embeddings. The decoder uses the cue embedding as the query and the position embedding as the key and value to perform a cross-attention operation. Subsequently, the decoder performs a third stage of reverse cross-attention operation on the position embeddings: in this stage, the position embedding is used as the query and the cue embedding is used as the key and value, and a cross-attention operation is performed again. The output encoded information is then input into the LayerNorm layer to normalize the encoded information, which helps stabilize the training process. After passing through the Linear layer, the feature dimensions are transformed. Finally, the Sigmoid function is used to introduce a non-linear transformation to enhance the expressive power of the model, and the final output is the accurate GU coverage predicted by the decoder.

[0068] 5.2) Workflow of the reverse network's mask decoder. The reverse network's mask decoder appends an output token to the cue token. This output token primarily outputs the optimal prediction positions for a variable number of ABS based on the number of ABS trained. The initial dimension of the output token is N. max *D(Invention N) max(The number of ABS is 1 to 5). After obtaining the location embedding and the concatenated cue token, both information are input into the mask decoder. Then, through a multi-attention mechanism, efficient feature interaction and fusion are performed to generate the optimal ABS location prediction based on the output token. The specific workflow can be divided into the following stages: The first stage is the self-attention stage of concatenating the cue token, where the mask decoder first performs a self-attention operation on the input cue token. The second stage is the cross-attention stage of the cue token and the location embedding, where the decoder uses the cue embedding as the query and the location embedding as the key and value, performing a cross-attention operation. Subsequently, the decoder performs a third stage of reverse cross-attention operation on the location embedding: in this stage, the location embedding is used as the query, and the cue embedding is used as the key and value, performing another cross-attention operation. Finally, based on the number of ABS N, the first N dimensions of the output tokens are taken as the final output token (dimension N*D), and the encoding is fed into the SAB layer. Each element of the final output token interacts with other elements in the set, thereby strengthening the correlation between the elements in the token set. Then, the final output token is input into the LayerNorm layer to normalize the tokens and stabilize the training process. After passing through the Linear layer, the feature dimensions are transformed. Finally, the Sigmoid function is used to introduce non-linear transformation to enhance the expressive power of the model. Finally, range scaling is performed to obtain the optimal ABS position prediction predicted by the model.

[0069] This invention constructs an ST-SAM model based on Set Transformer. First, the original image encoder is improved by introducing a Set Transformer to replace its original module with a position encoder. Second, the cue encoder is redesigned, introducing a PLE module to enable the model to encode input scalar parameters, such as ABS coverage radius, more reasonably and efficiently. Furthermore, a variable number of output tokens are introduced into the mask decoder, and the optimal position output for the variable number of ABS is achieved using the mask decoder's internal attention mechanism. Simultaneously, based on the SAM model architecture, this invention designs a forward-backward training network.

[0070] To verify the effectiveness of the method of the present invention, the present invention compares with GES, TD3, Kmeans, DBSCAN and the ST-SAM method proposed in this invention.

[0071] The ST-SAM network training parameters are as follows: the optimizer is set to the Adam optimizer, the initial learning rate is set to 0.001, the exponential decay rates of the first and second moments are set to 0.9 and 0.999 respectively, the weight decay rate is set to 1e-4, and the optimizer is dynamically adjusted using a dynamic learning rate. The network uses 500,000 training data points, 50,000 validation data points, and 50,000 test data points randomly generated. The map size is 256×256m. 2 The number of ABS ranges from 1 to 5, the coverage radius of ABS is 50 to 200m, the number of GU ranges from 100 to 500, and the GU locations are evenly scattered on the map. Figure 9 The visualization demonstrates the effectiveness of the ST-SAM model in predicting the location of a variable number (1–4) of Absorbed Units (ABS), where a–d correspond to the cases of 1–4 ABS, respectively. The visualization shows that as the number of ABS increases, the overall Guaranteed Unit (GU) coverage also significantly improves, gradually increasing from 0.32 in the case of 1 ABS to 0.82 in the case of 4 ABS. This illustrates that the model of this invention can adaptively select the optimal deployment location for a variable number of ABS.

[0072] Figure 10 This paper compares the coverage performance of different algorithms on the same map. In the graph, a represents the result of the GES exhaustive search method, b represents the result of the proposed ST-SAM method, c represents the TD3 result, d represents the K-means result, e represents the DBSCAN result, and f represents the result of the randomized method. It is clear from the graph that the ST-SAM model significantly outperforms the other methods, achieving a coverage rate of 0.47, which is close to the approximate optimal solution of 0.54 given by the GES algorithm. Although the GES method can provide an approximate optimal solution through exhaustive search, its computational cost is enormous, and in practical deployments it is often only used as an upper bound for theoretical performance. ST-SAM, on the other hand, achieves near-optimal performance with a much lower computational cost than GES.

[0073] Figure 11This paper analyzes the GU coverage performance of various algorithms under varying combinations of ABS (1-5) and GU (100-500) numbers. The subgraphs, from top to bottom, correspond to GU numbers ranging from 100 to 500. The results clearly show that, under different combinations of ABS and GU numbers, the GU coverage calculated by the proposed ST-SAM model is significantly higher than the other three comparison algorithms, including reinforcement learning TD3, K-means, and DBSCAN. With an ABS number of 1, the performance of the proposed model is very close to the theoretical upper bound obtained by GES. Enlarging the comparison to the average coverage of each algorithm when the ABS number is 2, although the problem becomes a combinatorial optimization problem, the coverage obtained by ST-SAM is still close to the upper bound obtained by GES. As the number of ABS increases, the GU coverage of the proposed model also increases, indicating that the model fully considers the joint coverage optimization, enabling multiple ABS to work together for coverage. This simulation experiment fully verifies the superiority and robustness of the ST-SAM model proposed in this invention under different ABS coverage numbers and different GU scales.

[0074] Figure 12 This analysis examines the GU coverage performance of various algorithms under varying ABS numbers (1–5) and coverage radii (50–200 m). The sub-graphs, from top to bottom, correspond to coverage radii of 50–100 m, 100–150 m, and 150–200 m, respectively. The results clearly show that, with different combinations of ABS and coverage radii, the GU coverage calculated by the proposed ST-SAM model is significantly higher than the other three comparative algorithms. With an ABS number of 1, the performance of the proposed model is very close to the theoretical upper bound obtained by GES. Furthermore, when comparing the average coverage of each algorithm with an ABS number of 2, the coverage obtained by ST-SAM remains close to the upper bound obtained by GES and is significantly higher than the other comparative algorithms. Moreover, as the ABS coverage radius increases, the algorithm's coverage also increases. When the coverage radius exceeds 150 m, the ST-SAM model can essentially cover all users. Therefore, this simulation experiment fully verifies the superiority and robustness of the ST-SAM model proposed in this invention under different ABS numbers and different ABS coverage radii.

[0075] A detailed analysis of the computation time of different algorithms in the ABS optimal position prediction problem is presented. The average ABS optimal position prediction time of each algorithm is shown in Table 1.

[0076] Table 1

[0077]

[0078] As shown in Table 1, the ST-SAM model proposed in this invention, by pre-embedding a variable number of output tokens in the mask decoder and employing a lightweight SetTransformer architecture, can efficiently predict the optimal position of a variable number of ABS in O(1) time complexity. In large-scale ABS coverage optimization problems, the ST-SAM model's solution speed is much slower than the GES algorithm, and its performance is close to the upper performance bound provided by the GES algorithm.

[0079] Table 2

[0080] algorithm GES ST-SAM finetune ST-SAM CR mean 0.881 0.824 0.796 Forecast time mean (s) 219 4.019 0.076

[0081] Furthermore, through further fine-tuning, the performance upper bound can be approximated again. The performance upper bound comparison after fine-tuning is shown in Table 2 (the test scenario is 2 ABS, 100 GU, and a GES step size of 4m). Therefore, the model proposed in this invention fully demonstrates the advantages of fast solution capability and efficient coverage performance.

[0082] This invention proposes a multi-ABS coverage optimization model based on SetTransformer (ST-SAM model). This model modifies the encoder and decoder of the SegmentAnythingModel (SAM) model to better suit the research problem. Utilizing its unique attention mechanism, the model encodes features of a variable number of ABS and GUs, effectively capturing the interaction relationships and spatial distribution characteristics between ABS and GUs. Furthermore, a forward-backward network training process is designed to further enhance the model's adaptability to complex conditions such as different numbers of ABS, coverage radii, and GU spatial distributions, enabling intelligent optimization selection of multi-ABS deployment locations.

[0083] The above embodiments are merely preferred embodiments of the present invention and should not be considered as limiting the scope of the present invention. All equivalent variations and improvements made within the scope of the present invention should still fall within the patent coverage of the present invention.

Claims

1. A multi-UAV joint coverage optimization method based on Set Transformer design, characterized in that... Includes the following steps: 1) System Model Construction and Problem Analysis: Define a wireless communication system containing multiple ABS and multiple GU, where the location set of GU is a known two-dimensional coordinate; based on the line-of-sight channel model and free space path loss model, set the ABS to fly at a given altitude, and define the signal-to-noise ratio threshold corresponding to its maximum ground coverage radius; define the coverage index of GU; model the multi-UAV joint coverage optimization problem. 2) Design the ST-SAM model. By introducing the Set Transformer module to build a position encoder, the PLE module to build a cue encoder, and introducing a variable number of output tokens and combining them with an internal attention mechanism to build a mask decoder, the SAM model is reconstructed to achieve encoding and decoding of relevant information and output of the optimal position of ABS. 3) Forward network training: Input randomly generated GU coordinates, ABS coordinates, ABS coverage radius and number of ABS; input the GU coordinates into the position encoder to generate position embedding, and input the ABS coverage radius, number and ABS position information processed by the position encoder into the prompt encoder to generate prompt embedding; The location embedding and cue embedding are input to the mask decoder, and the predicted GU coverage label is output. The network parameters are updated using the true coverage label under the current ABS placement condition as the label, through mean squared error loss and backpropagation. 4) Reverse network training: Input GU coordinates, ABS coverage radius and number; input GU coordinates into the position encoder to generate position embedding, input ABS coverage radius and number into the cue encoder to generate cue embedding; The maximum number of output tokens is initialized in the mask decoder, concatenated with the hint embedding, and then input into the decoder; the cross attention and multi-head self attention mechanism of the Transformer module are used for feature interaction and fusion, and the predicted optimal position coordinates of the ABS are output according to the number of target ABS. 5) Joint training of forward and backward networks: After the forward network is fully trained, all its parameters are frozen as a fixed coverage calculation module; when training the backward network, the ABS position coordinates output by the backward network are used to generate position codes through the frozen position encoder, and input into the prompt encoder of the forward network to calculate the corresponding GU coverage; with "all GUs are covered" as the ideal label, the error between the prediction result of the forward network and the ideal label is used to backpropagate and optimize the parameters of the backward network until the backward network stably outputs the optimal ABS position; 6) Output optimization results: Based on the trained ST-SAM model, input the target GU location, ABS number and coverage radius, and quickly output the optimal ABS prediction location to maximize GU coverage.

2. The multi-UAV joint coverage optimization method based on Set Transformer design as described in claim 1, characterized in that... In step 1), the specific steps for constructing the system model and analyzing the problem are as follows: Consider a wireless communication system with N ABS and M GU, where The location set of GU {q m } m∈M Given that q m ∈R 2×1 Let represent the two-dimensional coordinates of the m-th GU on the ground; assume that the ABS-GU communication channel is mainly dominated by the LoS link; also assume that the transmit power is fixed and the minimum signal-to-noise ratio required for reliable communication at the receiver is known; under the LoS model, the power gain of the ABS-GU channel follows the free-space path loss model, which is determined by the ABS-GU distance; to explore an algorithm that approximates the optimum in a finite time, assume that the ABS flies at a given altitude H, and its maximum coverage radius R on the ground corresponds to the SNR threshold; define the coverage index of GU m as: Where, p n ∈R 2×1 Represents the horizontal coordinate of ABSn, ||p n -q m || represents the distance between the projections of GU m and ABSn onto the ground; The goal of the multi-UAV joint coverage optimization problem is to maximize the coverage of the GU using N ABS (Automatic Protective Units). Therefore, the problem is modeled as (P1): Among them, C m q is the coverage metric for GU m. m ∈R 2×1 p represents the two-dimensional coordinates of the m-th GU on the ground. n ∈R 2×1 Represents the horizontal coordinate of ABSn, ||p n -q m || represents the distance between the projections of GU m and ABSn onto the ground, and R is the maximum coverage radius on the ground.

3. The multi-UAV joint coverage optimization method based on Set Transformer design as described in claim 1, characterized in that... In step 2), the ST-SAM model is designed by first improving the original image encoder by introducing SetTransformer to replace its original module with a position encoder. This position encoder encodes the GU map information to improve the model's ability to extract user location features and reduce computational resource consumption. The prompt encoder is redesigned by introducing a PLE module to enable the model to encode the input scalar parameters reasonably and efficiently. A variable number of output tokens are introduced into the mask decoder, and the optimal position output for the variable number of ABS is achieved by utilizing the attention mechanism inside the mask decoder.

4. The multi-UAV joint coverage optimization method based on Set Transformer design as described in claim 1, characterized in that... In step 3), the specific steps of the forward network training are as follows: In the forward network, the input data of the network includes randomly generated GU coordinates, ABS coordinates, ABS coverage radius R, and ABS quantity. The training label is the user coverage marker under the current ABS placement condition. The user coverage marker is a sequence containing only 0 / 1. First, the location coordinate information of the GU is input into the model's location encoder. The location encoder effectively encodes all GU locations on the map to form the corresponding location embedding, which is used as one of the inputs to the subsequent mask decoder. The ABS coverage radius and ABS quantity are used as cue inputs to the cue encoder for encoding to obtain the cue embedding. The ABS location information is also processed by the location encoder to obtain the ABS location embedding, which is then input into the cue encoder. The above location embedding and cue embedding are input together into the improved mask decoder. The predicted user coverage is output through the cross-attention and self-attention modules in the decoder mask. The corresponding label is the pre-calculated real user coverage obtained based on the current ABS placement. Then, the mean squared error loss between the two is calculated, and backpropagation is performed to update the network parameters.

5. The multi-UAV joint coverage optimization method based on Set Transformer design as described in claim 1, characterized in that... In step 4), the specific steps of training the reverse network are as follows: In the reverse network, the GU coordinate information is first input into the position encoder to obtain the corresponding GU position embedding; the coverage radius and number of ABS are used as prompts and input into the prompt encoder to obtain the corresponding prompt embedding. When initializing the mask decoder, a special output token for the maximum number of variable number of optimal UAV positions is pre-embedded. When the mask decoder receives the position embedding and cue embedding of the GU, it concatenates the output token with the cue embedding to obtain the concatenated output token, and then passes it into the improved mask decoder. The Transformer module within the decoder uses cross-attention and multi-head self-attention mechanisms to perform feature interaction and feature fusion between the concatenated output token and the GU position embedding; Cross-attention and multi-head self-attention mechanisms enable the model to simultaneously focus on the location information of the GU and the input cue information, accurately predicting and outputting the optimal ABS location coordinates as needed based on the required number of ABS outputs.

6. The multi-UAV joint coverage optimization method based on Set Transformer design as described in claim 5, characterized in that... The maximum number of output tokens is set to 5.

7. The multi-UAV joint coverage optimization method based on Set Transformer design as described in claim 1, characterized in that... In step 5), the specific steps of the joint training of the forward and reverse networks are as follows: the forward network, after preliminary training, fits an efficient calculation function, that is, calculates the GU coverage in real time based on the ABS position; when actually training the reverse network, all parameters of the pre-trained forward network are frozen as a fixed coverage calculation module; during the training process, the specific process of the reverse network is as follows: inputting the GU coordinate information into the position encoder to generate the GU position embedding, and simultaneously inputting the ABS coverage radius and ABS quantity information into the cue encoder to generate the cue embedding; the reverse network outputs the optimal ABS position coordinate prediction results through the improved mask decoder, and these position coordinates generate position encoding information through the frozen position encoder; This location encoding information is then input into the cue encoder in the forward network as part of the cue information, assisting the forward network in accurately calculating the GU coverage at the corresponding ABS location. Throughout the training process, the ideal label for GU coverage is set to all users being covered, i.e., the corresponding label max Coverage Flags sequence value is all set to 1, in order to maximize the ground user coverage rate. By calculating the error between the coverage rate predicted by the forward network and the ideal label, the parameters of the reverse network are optimized through backpropagation, and the optimization is continuously iterated. After the above repeated iterative training, the reverse network outputs the optimal ABS deployment location, achieving the maximization of ground GU coverage.

Images