Frequency spectrum situation map acquisition method and device, electronic equipment and storage medium

Abstract

The invention discloses a frequency spectrum situation map obtaining method and device, electronic equipment and a storage medium. The frequency spectrum situation map obtaining method comprises the steps of obtaining map structure features based on a sparse sampling point image and a building layout image of a target area; and based on a pre-trained TransUNet model, feature extraction is performed on the sparse sampling point image and the building layout image to obtain a first feature, the first feature and the graph structure feature are fused to obtain a fused feature, and a frequency spectrum situation map is obtained according to the fused feature. According to the frequency spectrum situation map obtaining method provided by the embodiment of the invention, the frequency spectrum situation map with relatively high precision can be obtained, less sampling data is used, and the generation effect of the frequency spectrum situation map can be improved and the physical interpretability of the model can be enhanced by fully utilizing the spatial sparse sampling points and the building layout image.

Description

Technical Field

[0001] The present application relates to the field of signal processing technology, and in particular to a method for acquiring a spectrum situation diagram, an acquisition device, an electronic device and a storage medium. Background Art

[0002] Electromagnetic spectrum is the basis of various wireless technologies such as communication, radar, navigation, and radio, and is also an important resource that is indispensable in the information age. With the rapid growth of wireless devices and technologies, electromagnetic spectrum resources are becoming increasingly valuable and limited. Electromagnetic spectrum situation awareness refers to obtaining real-time information on spectrum usage through dynamic monitoring and analysis of the electromagnetic spectrum, so as to grasp, analyze and predict the electromagnetic environment globally. Its core purpose is to timely and comprehensively understand the state of the electromagnetic spectrum and provide a basis for spectrum management and decision-making. Spectrum situation map generation is an important process to support electromagnetic spectrum situation awareness. It displays the state of spectrum usage and electromagnetic environment in a spatial and visual way. It not only reflects the signal strength or frequency band occupancy, but more importantly, it provides intuitive basic data for spectrum situation awareness, helping decision makers analyze and predict changes in the spectrum environment. In the face of the increasingly complex electromagnetic spectrum environment and the growing demand for frequency use, it is particularly urgent to maintain the order and security of the electromagnetic spectrum and improve the overall utilization efficiency of spectrum resources. At present, the regional demand for the construction of spectrum situation maps continues to expand to the wide-area space of land, sea, air and space. To this end, it is urgent to use limited sensing nodes to collect electromagnetic spectrum data and deeply mine electromagnetic environment information in wide-area geographic space. The related technology requires a lot of sampling data, the obtained spectrum situation map has low accuracy, and the generation effect of the spectrum situation map is not good.

[0003] The above statements are only used to provide background information related to the present application and do not necessarily constitute prior art. Summary of the invention

[0004] The purpose of the present application is to provide a method for acquiring a spectrum situation diagram, an acquisition device, an electronic device and a storage medium. In order to have a basic understanding of some aspects of the disclosed embodiments, a simple summary is given below. This summary is not a general review, nor is it intended to determine the key / important components or describe the scope of protection of these embodiments. Its only purpose is to present some concepts in a simple form as a preface to the detailed description that follows.

[0005] According to one aspect of an embodiment of the present application, a method for acquiring a spectrum situation map is provided, comprising:

[0006] Based on the sparse sampling point image and building layout image of the target area, the graph structure features are obtained;

[0007] Based on the pre-trained TransUNet model, perform feature extraction on the sparse sampling point image and the building layout image to obtain the first feature, fuse the first feature and the graph structure feature to obtain the fused feature, and obtain the spectral situation map according to the fused feature.

[0008] In some embodiments of the present application, obtaining the graph structure feature based on the sparse sampling point image and the building layout image of the target area includes:

[0009] Based on the pre-trained UNet model, the sparse sampling point image of the target area and the building layout image, obtain the emission source position;

[0010] Based on the pre-trained graph neural network, construct a graph structure according to the sparse sampling point image, the building layout image and the emission source position;

[0011] Extract the graph structure feature of the graph structure through the pre-trained graph neural network.

[0012] In some embodiments of the present application, obtaining the emission source position based on the pre-trained UNet model, the sparse sampling point image of the target area and the building layout image includes:

[0013] Use the pre-trained UNet model to process the sparse sampling point image and the building layout image to generate the construction result of the initial spectral situation map;

[0014] Obtain the emission source position according to the construction result of the initial spectral situation map.

[0015] In some embodiments of the present application, constructing a graph structure according to the sparse sampling point image, the building layout image and the emission source position based on the pre-trained graph neural network includes:

[0016] Mark the building areas in the building layout image, calculate the connected areas of the building layout image and return the building labels; the building layout image is a binary image representing the building distribution;

[0017] Extract the spatial positions and signal intensities of the non-zero values in the sparse sampling point image;

[0018] Through a pre-trained graph neural network, all node features, edge indices, and edge features are integrated into a graph-structured data. The node features include signal values, spatial coordinates, and attenuation factors, where the attenuation factor is determined by calculating the distance between the node and the position of the emission source and the length of the building traversed on the path. The edge index represents the connection relationship between nodes, and the edge features include the signal strength difference between two nodes and the distance difference from the two nodes to the emission source.

[0019] In some embodiments of the present application, based on a pre-trained TransUNet model, feature extraction is performed on the sparse sampling point image and the building layout image to obtain first features, including:

[0020] The first features are extracted from the sparse sampling point image and the building layout image through the convolutional blocks in the TransUNet model. The first features include signal features of different scales; each convolutional block includes a convolutional layer, a ReLU activation function, and a batch normalization layer.

[0021] In some embodiments of the present application, based on a pre-trained TransUNet model, the first features and graph-structured features are fused to obtain fused features, including:

[0022] The graph-structured features are mapped through the linear layer of the TransUNet model to the same number of channels as the bottleneck layer of the TransUNet model to obtain mapped features;

[0023] The mapped features are fused with the first features output by the convolutional block to obtain fused features.

[0024] In some embodiments of the present application, based on a pre-trained TransUNet model, a spectrum situation map is obtained according to the fused features, including:

[0025] The fused features are processed through the encoder of the TransUNet model to obtain second features;

[0026] The second features are mapped back to the original image size through the decoder to obtain a spectrum situation map.

[0027] In some embodiments of the present application, the process of obtaining second features by processing the fused features through the encoder of the TransUNet model includes:

[0028] Through the encoder of the TransUNet model, based on the self-attention mechanism to capture global information, the fused features are globally modeled and transformed to obtain signal features after global modeling, that is, second features.

[0029] According to another aspect of the embodiments of the present application, there is provided an apparatus for obtaining a spectrum situation map, including:

[0030] A graph structure feature acquisition module, configured to acquire graph structure features based on a sparse sampling point image and a building layout image of a target area;

[0031] A spectrum situation map acquisition module, configured to perform feature extraction on the sparse sampling point image and the building layout image based on a pre-trained TransUNet model to obtain first features, fuse the first features and the graph structure features to obtain fused features, and obtain a spectrum situation map according to the fused features.

[0032] According to another aspect of the embodiments of the present application, there is provided an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor, where the processor executes the computer program to implement the method for obtaining a spectrum situation map according to any embodiment of the present application.

[0033] According to another aspect of the embodiments of the present application, there is provided a computer-readable storage medium, on which a computer program is stored, and the computer program is executed by a processor to implement the method for obtaining a spectrum situation map according to any embodiment of the present application.

[0034] The technical solution provided by one aspect of the embodiments of the present application may include the following beneficial effects:

[0035] The method for obtaining a spectrum situation map provided by the embodiments of the present application acquires graph structure features based on a sparse sampling point image and a building layout image of a target area, performs feature extraction on the sparse sampling point image and the building layout image based on a pre-trained TransUNet model to obtain first features, fuses the first features and the graph structure features to obtain fused features, and obtains a spectrum situation map according to the fused features. In this way, a spectrum situation map with high accuracy can be obtained and less sampling data is used, and the generation effect of the spectrum situation map can be improved by making full use of spatial sparse sampling points and building layout images, and the physical interpretability of the model can be enhanced.

[0036] The above description is only an overview of the technical solutions of the embodiments of the present application. In order to be able to understand the technical means of the embodiments of the present application more clearly, it can be implemented according to the content of the description. And in order to make the above and other purposes, features and advantages of the embodiments of the present application more obvious and understandable, the specific embodiments of the present application are specifically exemplified below. Description of the Drawings

[0037] To more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings described below are only some embodiments recorded in the present application. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0038] Figure 1 The flowchart of the method for obtaining the spectrum situation diagram of an embodiment of the present application is shown.

[0039] Figure 2 The flowchart of obtaining the graph structure features based on the sparse sampling point image and the building layout image of the target area in an embodiment of the present application is shown.

[0040] Figure 3 The flowchart of obtaining the emitter position based on the pre-trained UNet model, the sparse sampling point image of the target area and the building layout image in an embodiment of the present application is shown.

[0041] Figure 4 The flowchart of constructing a graph structure based on the pre-trained graph neural network, the sparse sampling point image, the building layout image and the emitter position in an embodiment of the present application is shown.

[0042] Figure 5 The schematic structural diagram of the physics-driven enhanced network model based on TransUNet in an embodiment of the present application is shown.

[0043] Figure 6 The flowchart of creating a graph structure in an embodiment of the present application is shown.

[0044] Figure 7 The flowchart of the training process of the TransUNet model in an embodiment of the present application is shown.

[0045] Figure 8 The flowchart of the generation process of the spectrum situation diagram in an embodiment of the present application is shown.

[0046] Figure 9 The experimental result graph obtained by conducting an experiment on the method for obtaining the spectrum situation diagram of the embodiment of the present application is shown.

[0047] Figure 10 The structural block diagram of the device for obtaining the spectrum situation diagram of an embodiment of the present application is shown.

[0048] Figure 11 The structural block diagram of the electronic device of an embodiment of the present application is shown.

[0049] Figure 12Schematic diagram of a computer-readable storage medium according to an embodiment of the present application is shown. Detailed implementation manners

[0050] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be further described below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present application without creative efforts shall fall within the protection scope of the present application.

[0051] Those skilled in the art can understand that unless otherwise defined, all terms (including technical terms and scientific terms) used herein have the same meaning as the general understanding of those of ordinary skill in the technical field to which the present application belongs. It should also be understood that terms such as those defined in a general dictionary should be understood to have a meaning consistent with the meaning in the context of the prior art, and will not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0052] The goal of spectrum situation map construction is to obtain local electromagnetic environment data through sensing devices and use this data to estimate and model the global electromagnetic environment of the area of interest. The power spectral density (PSD) of a signal is one of the common ways to present the spectrum situation, which can reflect information such as the power, frequency occupancy, access protocol, and modulation method of the radiation source. The spatial loss function (SLF) is a presentation method in the spatial domain, which can reflect electromagnetic space characteristics and geographical characteristics such as spatial shadow fading, path loss, and obstacles. The core of the spectrum situation map construction task is how to recover the global electromagnetic environment information from limited sampling point data. Limited by the number of monitoring nodes, the movement path, and their detection, storage, calculation, and communication resources, there is still a problem of sparse spatial sampling in constructing a high-quality spectrum situation map.

[0053] Most of the spectrum situation generation methods in the related art focus on the generation methods in specific scenarios, and there is a lack of research on the representation and construction mechanism. Generally, the electromagnetic environment modeling methods are divided into two categories: parametric models and non-parametric models. The parametric model infers the key parameters required for the spectrum situation, such as SLF, PSD, position and other information, through known model assumptions. The advantage is that it can make full use of prior information and is suitable for scenarios with fewer observation points. The disadvantage is that when the model assumption does not match the actual scenario, the performance will be significantly affected. In contrast, the non-parametric model models the spectrum situation as a function dependent on the observed data and can handle scenarios without prior information. Common methods include interpolation, matrix / tensor completion, and neural network-based methods. In the method for constructing a spectrum situation map based on a neural network, the deep learning method based on a convolutional neural network has gradually become a research hotspot. These methods usually rely on a large amount of sampled data to ensure high-precision generation of the spectrum situation map. At the same time, these methods lack a reasonable explanation of the electromagnetic propagation model. At a low sampling rate, the neural network methods in the related art often produce large errors.

[0054] In view of the problems existing in the related art, an embodiment of the present application provides a method for obtaining a spectrum situation map. Based on the sparse sampling point image and the building layout image of the target area, the graph structure feature is obtained. Based on the pre-trained TransUNet model, the sparse sampling point image and the building layout image are subjected to feature extraction to obtain a first feature. The first feature and the graph structure feature are fused to obtain a fused feature, and a spectrum situation map is obtained according to the fused feature. In this way, a spectrum situation map with high precision can be obtained with less sampled data used. By making full use of the spatial sparse sampling points and the building layout image, the generation effect of the spectrum situation map can be improved, and the physical interpretability of the model can be enhanced.

[0055] Next, a method for obtaining a spectrum situation map, an obtaining device, an electronic device, and a storage medium proposed according to an embodiment of the present application will be described with reference to the accompanying drawings.

[0056] Reference Figure 1 As shown, an embodiment of the present application provides a method for obtaining a spectrum situation map. The method may include steps S10-S20:

[0057] S10. Based on the sparse sampling point image and the building layout image of the target area, obtain the graph structure feature.

[0058] In some embodiments, as shown in Figure 2 Based on the sparse sampling point image and the building layout image of the target area, obtaining the graph structure feature may include steps S101-S103:

[0059] S101. Based on the pre-trained UNet model, the sparse sampling point image of the target area, and the building layout image, obtain the emission source location.

[0060] Exemplarily, referring to Figure 3 As shown, based on the pre-trained UNet model, the sparse sampling point image of the target area, and the building layout image, obtaining the emission source location may include:

[0061] S1011. Use the pre-trained UNet model to process the sparse sampling point image and the building layout image to generate an initial spectrum situation map construction result;

[0062] S1012. According to the initial spectrum situation map construction result, obtain the emission source location. This emission source location may be the rough location of the emission source.

[0063] S102. Based on the pre-trained graph neural network, construct a graph structure according to the sparse sampling point image, the building layout image, and the emission source location.

[0064] Exemplarily, referring to Figure 4 As shown, based on the pre-trained graph neural network, constructing a graph structure according to the sparse sampling point image, the building layout image, and the emission source location may include:

[0065] S1021. Mark the building areas in the building layout image, calculate the connected regions of the building layout image, and return the building labels; the building layout image is a binary image representing the building distribution;

[0066] S1022. Extract the spatial positions and signal intensities of the non-zero values in the sparse sampling point image;

[0067] S1023. Through the pre-trained graph neural network, integrate all node features, edge indices, and edge features into a graph structure data.

[0068] Among them, the node features include signal values, spatial coordinates, and attenuation factors. The attenuation factor is determined by calculating the distance between the node and the emission source location and the length of the building traversed on the path. The edge index represents the connection relationship between nodes, and the edge features include the signal intensity difference between two nodes and the distance difference from the two nodes to the emission source.

[0069] S103. Extract the graph structure features of the graph structure through the pre-trained graph neural network.

[0070] S20. Based on the pre-trained TransUNet model, perform feature extraction on the sparse sampling point image and the building layout image to obtain a first feature, fuse the first feature and the graph structure feature to obtain a fused feature, and obtain a spectral situation map according to the fused feature.

[0071] In some embodiments, based on the pre-trained TransUNet model, performing feature extraction on the sparse sampling point image and the building layout image to obtain a first feature may include: S201. Extract a first feature from the sparse sampling point image and the building layout image through the convolutional blocks in the TransUNet model, and the first feature includes signal features of different scales. Each convolutional block includes a convolutional layer, a ReLU activation function, and a batch normalization layer.

[0072] Exemplarily, based on the pre-trained TransUNet model, fusing the first feature and the graph structure feature to obtain a fused feature may include:

[0073] S202. Map the graph structure feature through the linear layer of the TransUNet model to the same number of channels as the bottleneck layer of the TransUNet model to obtain a mapped feature;

[0074] S203. Fuse the mapped feature with the first feature output by the convolutional block to obtain a fused feature.

[0075] Exemplarily, based on the pre-trained TransUNet model, obtaining a spectral situation map according to the fused feature may include:

[0076] S204. Process the fused feature through the encoder of the TransUNet model to obtain a second feature.

[0077] Exemplarily, processing the fused feature through the encoder of the TransUNet model,

[0078] obtaining a second feature may include: Through the encoder of the TransUNet model, capture global information based on the self-attention mechanism, perform global modeling on the fused feature and transform the fused feature to obtain a signal feature after global modeling, that is, the second feature.

[0079] S205. Map the second feature back to the original image size through the decoder to obtain a spectral situation map.

[0080] In the embodiments of the present application, a dual-driven spectrum situation awareness data model based on a graph neural network (GNN) is realized. By combining spatially sparse sampling points and building layout information, a clear and dense spectrum situation map is generated. In the embodiments of the present application, a physics-driven enhanced network model based on TransUNet is adopted. This model includes a UNet model, a graph neural network (GNN), and a TransUNet model. The overall framework of this model is as Figure 5 shown. Specifically, in the embodiments of the present application, the graph neural network is used to combine the emission source location (obtained from the rough result generated by the UNet model), the sparse sampling point values in the sparse sampling point image, and the building layout image to construct a graph structure, and node features and edge features combined with physical information are extracted from this graph structure, fully considering the influence of the building layout on the propagation of radio signals, and improving the physical interpretability of the spectrum situation map generation method.

[0081] In addition, to address the problems of limited local receptive fields and insufficient modeling of long-range dependencies in the traditional UNet architecture for spectrum situation map generation tasks, the TransUNet architecture is introduced in the embodiments of the present application. By combining the U-Net architecture with a Transformer encoder, TransUNet can capture both local and global features simultaneously, thus making up for the deficiencies of the traditional UNet in modeling long-range dependencies. In the decoder part, TransUNet retains the structural advantages of U-Net, retaining both the local feature extraction ability and making full use of the Transformer's ability to model global context, improving the accuracy of spectrum situation map construction.

[0082] In the embodiments of the present application, a sparse sampling point enhancement strategy is also adopted. After generating the preliminary image, by calculating the second derivative values of the image, those sparse sampling points located in smooth regions and with high credibility are selected and added to the existing sparse sampling point image. By iterating this process multiple times, the number of sampling points with higher credibility is gradually increased, effectively improving the quality of the final spectrum situation map, especially in the case of extremely sparse sampling. Finally, the model framework proposed in the embodiments of the present application can generate a more accurate and dense spectrum situation map under sparse data conditions, significantly improving the quality of the generated image.

[0083] In a specific example, consider a spectrum situation map with a size of 256 * 256 pixels. Each pixel represents 1 m in the actual geographical space, and its value represents the path loss value after a certain processing. Therefore, this spectrum situation map can represent the electromagnetic spectrum usage in a geographical space of 256 m * 256 m. Specifically, the value of each pixel in this image first undergoes a truncation process, that is, extremely low path loss values are first removed to highlight the importance of high path gain. Secondly, the range of path loss values in the dataset is linearly scaled to the grayscale value range of 0 - 255. Higher path gain (low path loss) corresponds to the bright area with a larger grayscale value in the image, and lower path gain (high path loss) corresponds to the dark area with a smaller grayscale value in the image. The specific linear transformation is where PL(d) is the path loss value at distance d, and max(PL) and min(PL) are the maximum and minimum path loss values in the image respectively.

[0084] In the method of the embodiment of the present application, a physical model driven based on a graph neural network is realized. The creation process of the graph structure in the embodiment of the present application focuses on how to integrate the physical characteristics of signal propagation and the influence of buildings into the graph neural network model. Specifically, in the embodiment of the present application, the creation process of the graph structure transforms physical factors such as signal strength, spatial position, building occlusion effect, and path attenuation into node and edge features in the graph, enabling the graph neural network to embody physical drive in learning.

[0085] Exemplarily, in the process of creating the graph structure, the input building image is a binary image representing the building distribution. First, the building area is marked, where the building area is marked as 1 and the non - building area is marked as 0. By calculating the connected regions of the building layout image and returning the building labels, a basis can be established for the subsequent relationship between nodes and the building area. The marking of the building area directly reflects the influence of buildings on signal attenuation during signal propagation. By introducing the label information of the building area into the graph structure, the subsequent graph neural network can learn the influence of buildings on signal propagation.

[0086] Secondly, extract the spatial positions and signal intensities of non - zero values in the sparse sampling point image, where the signal intensity S of node i i represents the physical signal value of a spatial position, and the spatial position of the node [x i , y ireflects the distribution in the physical space. These node features provide information about the spatial distribution and signal strength for the subsequent graph neural network model, enabling the model to learn how to update the node features according to the physical space and signal propagation characteristics. In addition to the signal value and coordinates, the features of each node not only include the signal value and spatial coordinates but also an attenuation factor, which reflects the influence of the building during signal propagation. The attenuation factor is calculated based on the distance between the node and the emission source and the length of the building area traversed in the node's path, and can simulate the shielding or weakening effect of the building on signal propagation. Usually, the signal attenuates when passing through the building, and the calculation formula is:

[0087] Attenuation Factor(A i )=exp(-β·L buildings )

[0088] Where, A i is the attenuation factor of node i, β is the path loss factor, taking 0.1, L buildings is the length of the building area passed by the path between node i and the emission source. The calculation of the attenuation factor directly reflects the physical mechanism in signal propagation. By incorporating the attenuation factor into the node features, the network can model signal attenuation in the subsequent learning process and fully consider the influence of the building on the signal. To sum up, the node features include the signal value, coordinates, and attenuation factor. The features of each node can be represented as: h i =[S i ,x i ,y i ,A i .

[0089] Exemplarily, in the graph structure, the relationship between nodes is represented by edges. Nodes are grouped according to the building labels they are in, ensuring that edge connections are established between nodes within the same building group, that is, only nodes that pass through exactly the same buildings from the emission source will establish edge connections. Since in addition to the influence of the building, the propagation of the signal is also affected by the distance. Therefore, the edge features include the signal difference and the distance difference, and their calculation is based on the attenuation and distance effects in physical signal propagation, which can help the model better understand the relationship between nodes and the actual physical process of signal propagation. Specifically, the signal difference is obtained by calculating the signal strength difference between two nodes, reflecting the attenuation during signal propagation. The calculation formula is:

[0090] ΔS ij =|S i -S j |

[0091] Where, S i and S jThey are the signal strengths of node i and node j respectively.

[0092] The distance difference is obtained by calculating the distance differences between two nodes and the transmitter, and can further simulate the signal changes caused by distance changes during signal propagation. Specifically, the distance from each node to the transmitter is calculated by the Euclidean distance, and this distance affects the signal strength and propagation. The formula is as follows:

[0093] D u =||p u -p transmitter || 2

[0094] where p i and p transmitter are the spatial coordinates of node i and the transmitter respectively. The distance difference is used to describe the distance differences between nodes and affects the degree of signal attenuation. The formula is as follows:

[0095] ΔD ij =|D i -D j |

[0096] where D i and D j are the distances from node i and node j to the transmitter respectively.

[0097] Finally, all node features, edge indices, and edge features are integrated into a graph structure data. Among them, node features include signal values, spatial coordinates, and attenuation factors; edge indices represent the connection relationships between nodes; edge features include information such as signal differences and distance differences, and are expressed as e ij =[ΔS ij ,ΔD ij . This graph data structure provides the input for the subsequent graph neural network model, enabling the network to learn based on characteristics such as attenuation, distance, and building influence in the physical model. The overall process of creating the graph structure is as Figure 6 shown.

[0098] Exemplarily, after creating the above graph structure, data will be input into the graph neural network model for feature extraction and signal propagation modeling. During the training process of the graph neural network, the model processes node features and edge features through the TransformerConv graph convolutional layer to learn the relationships between nodes in the graph and their physical properties. Different from traditional graph convolutional networks (GCNs), TransformerConv incorporates the self-attention mechanism in Transformer, enabling it to consider more complex context relationships and the influence of edge features during node feature aggregation. It has significant advantages in capturing long-range dependencies between nodes and the influence of edge features, and can more accurately simulate complex relationships in the physical environment. Each layer of convolutional operation not only considers the topological structure between nodes, but also reflects physical effects such as signal propagation attenuation, path length, and building occlusion through edge features. The node features are gradually updated through multiple layers of TransformerConv and residual connections. The output of each convolutional layer will be used as the input for the next layer of convolution, enabling the model to learn complex signal propagation laws through multi-level graph convolutions. In TransformerConv convolution, the features of a node are updated based on the information of neighboring nodes, and the calculation formula is:

[0099]

[0100] where h i ′ is the updated feature of node i, h j is the feature of node j, j ∈ N(u) is the set of neighboring nodes connected to node i, c ij is the normalization coefficient, W is the weight matrix of the graph convolution, and b is the bias term.

[0101] Exemplarily, in each layer of convolution, TransformerConv uses the self-attention mechanism to perform weighted aggregation on node features, considering the importance of each neighboring node in feature update. TransformerConv does not simply update node features by summing neighboring nodes, but adaptively adjusts the relative influence between nodes through weights. This self-attention mechanism helps the model effectively balance between the local structure and global information of the graph, and can capture more complex signal propagation laws. In addition, residual connections ensure that information can propagate smoothly between layers, avoid the problem of gradient vanishing, and accelerate the training process of the model. The updated node features will be further processed through a multi-layer perceptron (MLP) to form the final node representation, and the calculation formula is:

[0102] h′ i = MLP(h i + h′ i )

[0103] Among them, h i is the initial feature of node i, and h' i is the feature after being updated by graph convolution. The MLP layer is used to further optimize the node representation. During the learning process, the graph neural network will adjust the transmitted information between nodes according to physical characteristics (such as signal attenuation factor and propagation distance). Through training, the network can adaptively learn the specific impacts of factors such as buildings and distances on signal propagation, so as to more accurately simulate the signal propagation situation in the actual environment.

[0104] In the embodiment of the present application, by embedding physical characteristics (such as building influence, distance, signal attenuation, etc.) into the graph structure and combining with the graph neural network model for learning and feature extraction, physical-driven modeling of the signal propagation process is realized, which can not only capture local and global relationships in the graph structure, but also automatically adjust model parameters related to physical characteristics during the learning process.

[0105] In the method of the embodiment of the present application, a physical-driven enhanced network model based on TransUNet is adopted. In the embodiment of the present application, a physical-driven enhanced network model that combines the local feature extraction ability of UNet and the global modeling ability of Transformer and adds a graph neural network (GNN) module is used to extract physical propagation features from sparse signal images and building information. The overall structure of the model combines the advantages of convolutional neural network (CNN), graph neural network (GNN) and Transformer encoder, and can effectively capture physical characteristics in the signal propagation process, such as the occlusion effect of buildings, signal attenuation and changes in signal propagation paths.

[0106] Exemplarily, the encoder part of the model is responsible for extracting high-level features from the input data. The input includes sparse signal images and building information images, and these images are concatenated as the initial input of the model. First, the input is processed through three convolutional blocks. Each convolutional block consists of a series of convolutional layers, ReLU activation functions and batch normalization (BatchNorm) layers to extract signal features at different scales. After each convolutional block, a max pooling layer is used for spatial dimensionality reduction to gradually reduce the spatial dimension of the image while extracting more abstract features. In the bottleneck part of the network, deeper convolutional layers are used to perform more in-depth feature learning on the signal to further improve the feature extraction ability of the model. This part performs two convolutional operations to keep the spatial size of the feature map unchanged while extracting higher-level feature representations.

[0107] Exemplarily, the graph neural network module is used to combine the graph structure with physical characteristics in signal propagation (such as the occlusion effect of buildings, signal attenuation, etc.), further optimizing the feature learning process. By inputting the estimated transmitter location, sparse sampling point images, and building information, a graph structure-based model is constructed. The features of each node include not only signal strength and spatial coordinates but also an attenuation factor related to building occlusion and signal propagation paths. This attenuation factor is determined by calculating the distance between the node and the transmitter and the length of the buildings traversed on the path, and can accurately simulate the weakening effect of buildings on signals. In the graph neural network module, the node features are processed by TransformerConv, and at the same time, edge features (such as signal difference, distance difference, etc.) are used for node information transmission. Through the aggregation of all node features, the graph neural network module finally outputs a feature representation modeled with physical characteristics.

[0108] Exemplarily, the features output from the GNN module are mapped to the same number of channels as the bottleneck layer through a linear layer and fused with the features from the convolutional encoder. Specifically, the features extracted by the GNN are expanded to the same spatial dimension as the bottleneck features and weighted-fused with the output of the bottleneck layer through a residual connection, thus fully combining the feature information from the GNN and the convolutional network. This process ensures that the influence of physical characteristics on signal propagation is effectively integrated into the feature representation of the entire model. The formula is:

[0109] x fused = x conv + α·x gnn

[0110] where x conv and x gnn are the features extracted by the convolutional encoder and the graph neural network respectively, and α is the weight coefficient. After the output of the bottleneck layer fused with the GNN features, the signal features are further processed by the Transformer encoder. The role of the Transformer encoder is to perform global modeling on the features, capture global information through the self-attention mechanism, enabling the model to effectively capture long-range dependencies in the signal propagation process, and transform the input features through multiple encoding layers to output the signal features after global modeling. This process is expressed as:

[0111] h output = TransformerEncoder(h 1 , h 2 , …, h N )

[0112] where h 1 , h 2 , …, hN is the input feature, h output is the feature after global modeling. In the Transformer encoder part, in order to ensure that the model can process information at different positions in the sequence, positional encoding is adopted, which is generated by sine and cosine functions, and the formula is:

[0113]

[0114] where pos is the position index, i is the dimension index, and d model is the model dimension.

[0115] Exemplarily, the decoder part is used to map the features processed by the Transformer encoder back to the original image size. The features are upsampled through three transposed convolutional layers to gradually restore the spatial resolution, and the features at different levels are further fused through convolutional blocks. At each step of upsampling, the features of the previous layer are concatenated with the features of the corresponding encoding layer to ensure that the detailed information at the low level can be effectively restored. At the end of the decoder, the output layer maps the restored features to a single-channel output through a 1x1 convolution to obtain the final prediction result.

[0116] The model of the embodiment of the present application can effectively model the physical mechanism in the signal propagation process by combining the advantages of GNN, Transformer, and CNN. The GNN part models the signal attenuation driven by physics, considering factors such as building occlusion and distance effect, and can effectively capture the complex relationships between different spatial regions, improving physical interpretability; the Transformer part captures the long-distance dependencies in signal propagation through global modeling; while the convolutional network captures the local features of the signal at the low level to ensure that the model can gradually extract the multi-dimensional information of the signal from details to the global. Finally, the model realizes the accurate prediction and optimization of the complex signal propagation process through multi-scale feature fusion and physical property modeling.

[0117] In the embodiment of the present application, sparse sampling enhancement based on the second derivative value is adopted. In the process of generating the spectral situation map by using the physics-driven enhanced network model based on TransUNet, the sparse sampling point image and building information are provided to the model as prior information. Experiments prove that when the size of the spectral situation map (actual geographical area) is fixed, increasing the number of spatial sparse sampling points in the generation stage is expected to improve the quality of the generated image. Therefore, the embodiment of the present application proposes to improve the quality of the generated image by increasing the sparse sampling points multiple times after using the above model to generate the preliminary result.

[0118] At the numerical level, the path loss value undergoes a linear transformation, but it still maintains a logarithmic form in terms of mathematical description and growth pattern. Therefore, in areas without building mutations, especially in areas far from the signal source, the growth of the logarithmic function gradually stabilizes. The second derivative value can represent the rate of change of the path loss change rate of each pixel point in the spectral situation map. It may be slightly larger near the signal source, but it will quickly tend to 0 as the distance increases, indicating that the gray-scale change gradually becomes stable. Therefore, the second-order gradient is close to 0 in most areas of the spectral situation map, especially in areas without mutations or significant changes. In terms of calculation, first, a 3×3 Laplacian operator is defined to calculate the second derivative. The specific convolution kernel used is: Then, the defined Laplacian operator is used to convolve each pixel point of the generated prediction image to obtain the second derivative value of each pixel point in the entire image. Since the building image is a pure black area in the spectral situation map, the pixels in the building area are excluded when calculating the second derivative.

[0119] In the embodiment of the present application, a certain second-order gradient loss threshold M is set as a constraint to select new sparse sampling points. In the generated spectral situation map, the second derivative values in relatively smooth areas are relatively small. Therefore, by setting a reasonable upper limit of the second-order gradient threshold, a certain number of new sampling points can be selected from the relatively smooth areas of the image. It should be noted that when screening new sparse sampling points, it is generally carried out outside the existing sparse sampling points and the building area, and the sparse sampling point values that meet the conditions are added to the existing sparse sampling point image to improve the quality of the generated image.

[0120] In a specific example, the training process of the TransUNet model is as Figure 7 shown, and the generation process of the spectral situation map in this example is as Figure 8 shown.

[0121] The method for obtaining the spectral situation map according to the embodiment of the present application is experimentally tested on the DPM simulation method dataset of the RadioMapSeer dataset. Referring to Figure 9 shown, where Ground Truth means the reference true value. When the prior condition is 50 sampling points, the average RMSE of the method proposed in the embodiment of the present application is lower than 0.018, the NMSE is lower than 0.003, and the PSNR exceeds 35dB. The effect of the spectral situation map generated by the method proposed in the embodiment of the present application is better and the accuracy is higher.

[0122] In the embodiments of the present application, through the dual-drive spectrum situation map cognition based on the graph neural network data model, a graph structure is constructed using the emitter location, sparse sampling points, and building layout information, combined with the physical propagation attenuation factor, and fully considering the occlusion effect of buildings on signal propagation, enhancing the physical interpretability of this task.

[0123] In the embodiments of the present application, the TransUNet model is introduced into the spectrum situation map generation task, and the advantages of TransUNet in local and global feature extraction are fully utilized in the dual-drive architecture of the data model, overcoming the local receptive field limitation of the traditional UNet, thereby enhancing the reconstruction accuracy of the spectrum situation map.

[0124] In the embodiments of the present application, a sparse sampling point enhancement strategy based on the second derivative constraint is adopted, and highly credible sparse sampling points are added multiple times during the image generation process, improving the accuracy of the final spectrum situation map.

[0125] Aiming at the problem that the parametric model method requires prior information such as known channel propagation characteristics and depends on estimating the key parameters required to generate the spectrum situation map, the embodiments of the present application propose a spectrum situation cognition method based on the graph neural network and TransUNet, which can directly generate a global spectrum situation map according to the spatial sparse sampling point values and building layout information without relying on prior parameters such as the path loss exponent, emitter location, and emitter power.

[0126] Aiming at the problems that non-parametric models such as the interpolation method, low-rank completion method, and traditional neural network method require high spatial sampling and their performance is greatly affected by the sampling point position, the method of the embodiments of the present application uses dual-drive spectrum situation cognition through the data model, fully utilizes the graph neural network, and the features of each node consider the signal value, position coordinates of the sampling point, and signal attenuation caused by buildings, and edge connections are established between the nodes in the same building area, taking the signal intensity difference and distance difference as edge features. By considering physical characteristics such as building layout and signal attenuation, the quality of the generated image can be improved at an extremely low spatial sampling rate, and the dependence on the sampling point position and quantity can be reduced.

[0127] Based on the graph neural network and TransUNet architecture, combined with the sparse sampling enhancement strategy, the embodiments of the present application construct a spectrum situation map, which can effectively integrate sparse sampling point and building layout information and directly generate a high-precision spectrum situation map without relying on a channel propagation model or other prior knowledge.

[0128] The spectrum situation map acquisition method proposed in the embodiments of the present application can generate a dense and clear spectrum situation map at a low spatial sampling rate, is not easily affected by the sampling point position, and the quality of the generated image is better than the interpolation, low-rank completion, and neural network-based baseline methods in the related technology.

[0129] In the embodiments of the present application, through the data model dual-driven spectrum situation awareness based on the graph neural network (GNN), by making full use of the spatial sparse sampling points and the building layout images, the generation effect of the spectrum situation map is improved, and at the same time, the physical interpretability of the model is enhanced. Specifically, the method of the embodiments of the present application first uses the basic UNet architecture to process the sparse sampling point data and the building layout images, generates a preliminary rough result of the spectrum situation map construction, and preliminarily estimates the emission source position through the pixel value maximization strategy. Subsequently, based on the graph neural network, a graph structure is constructed from the sparse sampling point image, the building layout image, and the estimated emission source position, and the graph structure features are extracted. At the same time, the TransUNet model is used to extract features from the sparse sampling point image and the building layout image, and the extracted features are fused with the features extracted by the graph neural network, and a more accurate spectrum situation map is generated. In addition, in the embodiments of the present application, through the sparse sampling point enhancement method, more reliable new sparse sampling points are selected from the generated spectrum situation map through reasonable constraint conditions (the second derivative value in the image is less than a certain threshold), further improving the quality of the generated map. The simulation results show that the proposed data model dual-driven method has a significant performance improvement under low sampling rate conditions compared with the interpolation method and the existing neural network baseline scheme in the related art.

[0130] The descriptions of the above embodiments tend to emphasize the differences between the embodiments. The same or similar parts can be referred to each other. For the sake of brevity, they will not be repeated herein.

[0131] Referring Figure 10 As shown, another embodiment of the present application provides an apparatus for obtaining a spectrum situation map, which may include:

[0132] A graph structure feature acquisition module, configured to acquire graph structure features based on the sparse sampling point image and the building layout image of the target area;

[0133] A spectrum situation map acquisition module, configured to extract first features from the sparse sampling point image and the building layout image based on a pre-trained TransUNet model, fuse the first features and the graph structure features to obtain fused features, and obtain a spectrum situation map according to the fused features.

[0134] Exemplarily, the graph structure feature acquisition module includes:

[0135] An emission source position acquisition unit, configured to acquire an emission source position based on a pre-trained UNet model, the sparse sampling point image and the building layout image of the target area;

[0136] A graph structure construction unit, configured to construct a graph structure based on a pre-trained graph neural network according to the sparse sampling point image, the building layout image, and the emission source location;

[0137] A graph structure feature extraction unit, configured to extract graph structure features of the graph structure through the pre-trained graph neural network.

[0138] Exemplarily, the emission source location acquisition unit includes:

[0139] An initial spectrum situation map construction result generation subunit, configured to process the sparse sampling point image and the building layout image by using a pre-trained UNet model to generate an initial spectrum situation map construction result;

[0140] An emission source location acquisition subunit, configured to acquire the emission source location according to the initial spectrum situation map construction result.

[0141] Exemplarily, the graph structure construction unit is further specifically configured to:

[0142] Mark the building areas in the building layout image, calculate the connected regions of the building layout image and return building labels; the building layout image is a binary image representing the building distribution;

[0143] Extract the spatial positions and signal intensities of non-zero values in the sparse sampling point image;

[0144] Through the pre-trained graph neural network, integrate all node features, edge indices, and edge features into a graph structure data, where the node features include signal values, spatial coordinates, and attenuation factors, the attenuation factor is determined by calculating the distance between the node and the emission source location and the length of the building traversed on the path, the edge index represents the connection relationship between nodes, and the edge features include the signal intensity difference between two nodes and the distance difference from the two nodes to the emission source.

[0145] The spectrum situation map acquisition module includes a first feature extraction unit, a fusion unit, and a spectrum situation map acquisition unit. The first feature extraction unit is configured to perform feature extraction on the sparse sampling point image and the building layout image based on a pre-trained TransUNet model to obtain first features. The fusion unit is configured to fuse the first features and the graph structure features based on a pre-trained TransUNet model to obtain fused features. The spectrum situation map acquisition unit is configured to obtain a spectrum situation map based on the fused features according to a pre-trained TransUNet model.

[0146] Exemplarily, the first feature extraction unit is further specifically configured to extract a first feature from the sparse sampling point image and the building layout image through a convolutional block in the TransUNet model, where the first feature includes signal features of different scales; each convolutional block includes a convolutional layer, a ReLU activation function, and a batch normalization layer.

[0147] Exemplarily, the fusion unit is further specifically configured to: map the graph structure feature to the same number of channels as the bottleneck layer of the TransUNet model through a linear layer of the TransUNet model to obtain a mapped feature; fuse the mapped feature with the first feature output by the convolutional block to obtain a fused feature.

[0148] Exemplarily, the spectral situation map acquisition unit is further specifically configured to: process the fused feature through the encoder of the TransUNet model to obtain a second feature; map the second feature back to the original image size through the decoder to obtain a spectral situation map.

[0149] Exemplarily, the process of processing the fused feature through the encoder of the TransUNet model to obtain a second feature includes: through the encoder of the TransUNet model, capturing global information based on the self-attention mechanism, globally modeling the fused feature and transforming the fused feature to obtain a globally modeled signal feature, that is, the second feature.

[0150] The spectral situation map acquisition device provided by the embodiments of the present application obtains a graph structure feature based on a sparse sampling point image and a building layout image of a target area, extracts a first feature from the sparse sampling point image and the building layout image based on a pre-trained TransUNet model, fuses the first feature and the graph structure feature to obtain a fused feature, and obtains a spectral situation map according to the fused feature. In this way, a spectral situation map with high accuracy can be obtained and less sampling data is used. By making full use of spatial sparse sampling points and building layout images, the generation effect of the spectral situation map can be improved, and the physical interpretability of the model can be enhanced.

[0151] The descriptions of the above embodiments tend to emphasize the differences between the embodiments. The same or similar parts can be referred to each other. For the sake of brevity, they will not be repeated herein.

[0152] Another embodiment of the present application provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. The processor executes the computer program to implement the method of any of the above embodiments.

[0153] ReferenceFigure 11 As shown in Figure 11 , the electronic device 10 may include: a processor 100, a memory 101, a bus 102, and a communication interface 103. The processor 100, the communication interface 103, and the memory 101 are connected through the bus 102. A computer program that can run on the processor 100 is stored in the memory 101. When the processor 100 runs this computer program, it executes the method provided in any of the foregoing embodiments of the present application.

[0154] Among them, the memory 101 may include a high-speed random access memory (RAM: Random Access Memory), and may also include a non-volatile memory, such as at least one disk memory. Through at least one communication interface 103 (which may be wired or wireless), a communication connection is realized between this device network element and at least one other network element, and the Internet, a wide area network, a local area network, a metropolitan area network, etc. can be used.

[0155] The bus 102 may be an ISA bus, a PCI bus, an EISA bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. Among them, the memory 101 is used to store a program. After receiving an execution instruction, the processor 100 executes this program. The method disclosed in any of the foregoing embodiments of the present application can be applied to the processor 100 or implemented by the processor 100.

[0156] The processor 100 may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method can be completed by the integrated logic circuit in the hardware of the processor 100 or an instruction in software form. The above-mentioned processor 100 may be a general-purpose processor, which may include a central processing unit (CPU for short), a network processor (NP for short), etc.; it may also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components. It can implement or execute the various methods, steps, and logic block diagrams disclosed in the embodiments of the present application. The general-purpose processor may be a microprocessor or this processor may also be any conventional processor, etc. The steps of the method disclosed in combination with the embodiments of the present application can be directly embodied as being executed and completed by a hardware decoding processor, or executed and completed by a combination of hardware and software modules in the decoding processor. The software module may be located in a mature storage medium in the art such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory, or an electrically erasable programmable memory, a register, etc. This storage medium is located in the memory 101, and the processor 100 reads the information in the memory 101 and combines its hardware to complete the steps of the above method.

[0157] The electronic device provided by the embodiment of the present application and the method provided by the embodiment of the present application are based on the same inventive concept and have the same beneficial effects as the method adopted, run or implemented by it.

[0158] Another embodiment of the present application provides a computer-readable storage medium, on which a computer program is stored, and the computer program is executed by a processor to implement the method of any of the above embodiments. Refer to Figure 12 As shown, the computer-readable storage medium shown is an optical disc 20, on which a computer program (i.e., a program product) is stored. When the computer program is run by a processor, it will execute the method provided by any of the foregoing embodiments.

[0159] It should be noted that examples of computer-readable storage media may also include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory, or other optical and magnetic storage media, which will not be elaborated here one by one.

[0160] The computer-readable storage medium provided by the above embodiment of the present application and the method provided by the embodiment of the present application are based on the same inventive concept and have the same beneficial effects as the method adopted, run or implemented by the application program stored in it.

[0161] It should be noted that:

[0162] The term "module" is not intended to be limited to a specific physical form. Depending on the specific application, a module can be implemented as hardware, firmware, software, and / or a combination thereof. In addition, different modules can share common components or even be implemented by the same components. There may or may not be a clear boundary between different modules.

[0163] The algorithms and displays provided herein are not inherently related to any particular computer, virtual device, or other equipment. Various general-purpose devices can also be used in conjunction with the examples based herein. Based on the above description, the structure required to construct such devices is obvious. In addition, the present application is not directed to any specific programming language. It should be understood that the content of the present application described herein can be implemented using various programming languages, and the description of a specific language above is to disclose the best implementation mode of the present application.

[0164] It should be understood that although the steps in the flowcharts of the accompanying drawings are shown sequentially as indicated by the arrows, these steps are not necessarily executed sequentially in the order indicated by the arrows. Unless there is a clear indication in this document, there is no strict order restriction for the execution of these steps, and they can be executed in other orders. Moreover, at least a part of the steps in the flowcharts of the accompanying drawings may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same moment, but can be executed at different moments, and their execution order is not necessarily sequential, but can be executed alternately or in turn with at least a part of other steps or sub-steps or stages of other steps.

[0165] The above embodiments only express the implementation manners of the present application, and the description thereof is relatively specific and detailed, but it should not be construed as a limitation on the patent scope of the present application. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can still be made, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the appended claims.

Claims

1. A method for obtaining a spectrum situation diagram, characterized in that: include: Based on the sparse sampling point image and building layout image of the target area, the graph structure features are obtained; Based on the pre-trained TransUNet model, feature extraction is performed on the sparse sampling point image and the building layout image to obtain a first feature, the first feature and the graph structure feature are fused to obtain a fused feature, and a spectrum situation map is obtained according to the fused feature.

2. The method according to claim 1, characterized in that The step of obtaining the graph structure features based on the sparse sampling point image and the building layout image of the target area includes: Obtaining the location of the emission source based on the pre-trained UNet model, the sparse sampling point image of the target area, and the building layout image; Based on a pre-trained graph neural network, a graph structure is constructed according to the sparse sampling point image, the building layout image and the emission source position; The graph structure features of the graph structure are extracted through the pre-trained graph neural network.

3. The method according to claim 2, characterized in that The method of obtaining the emission source position based on the pre-trained UNet model, the sparse sampling point image of the target area and the building layout image includes: Using a pre-trained UNet model to process the sparse sampling point image and the building layout image to generate an initial spectrum situation map construction result; According to the initial spectrum situation map construction result, the emission source position is obtained.

4. The method according to claim 2, characterized in that: The pre-trained graph neural network constructs a graph structure according to the sparse sampling point image, the building layout image and the emission source position, including: Marking the building areas in the building layout image, calculating the connected areas of the building layout image and returning the building labels; the building layout image is a binary image representing the distribution of buildings; Extracting the spatial position and signal strength of non-zero values ​​in the sparse sampling point image; Through the pre-trained graph neural network, all node features, edge indexes and edge features are integrated into a graph structure data. The node features include signal value, spatial coordinates and attenuation factor. The attenuation factor is determined by calculating the distance between the node and the location of the emission source and the length of the building traversed on the path. The edge index represents the connection relationship between the nodes, and the edge features include the signal strength difference between the two nodes and the distance difference from the two nodes to the emission source.

5. The method according to claim 1, characterized in that Based on the pre-trained TransUNet model, feature extraction is performed on the sparse sampling point image and the building layout image to obtain a first feature, including: A first feature is extracted from the sparse sampling point image and the building layout image through a convolution block in the TransUNet model, wherein the first feature includes signal features of different scales; each of the convolution blocks includes a convolution layer, a ReLU activation function, and a batch normalization layer.

6. The method according to claim 5, characterized in that Based on the pre-trained TransUNet model, the first feature and the graph structure feature are fused to obtain the fused features, including: Mapping the graph structure features to the same number of channels as the bottleneck layer of the TransUNet model through the linear layer of the TransUNet model to obtain mapped features; The mapped feature is fused with the first feature output by the convolution block to obtain a fused feature.

7. The method according to claim 1, characterized in that Based on the pre-trained TransUNet model, a spectrum situation map is obtained according to the fused features, including: Processing the fused features through the encoder of the TransUNet model to obtain a second feature; The second feature is mapped back to the original image size through the decoder to obtain a spectrum situation map.

8. The method according to claim 7, characterized in that Processing the fused features through the encoder of the TransUNet model to obtain a second feature includes: Through the encoder of the TransUNet model, global information is captured based on the self-attention mechanism, the fused features are globally modeled and transformed, and the signal features after global modeling, i.e., the second features, are obtained.

9. A device for acquiring a spectrum situation diagram, characterized in that: include: A graph structure feature acquisition module, used to acquire graph structure features based on sparse sampling point images and building layout images of the target area; The spectrum situation map acquisition module is used to extract features from the sparse sampling point image and the building layout image based on a pre-trained TransUNet model to obtain a first feature, fuse the first feature with a graph structure feature to obtain a fused feature, and obtain a spectrum situation map according to the fused feature.

10. An electronic device, characterized in that: It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method for acquiring a spectrum situation diagram as described in any one of claims 1 to 8.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that: The computer program is executed by a processor to implement the method for acquiring a spectrum situation map as described in any one of claims 1-8.

Images