Method for constructing electromagnetic spectrum map based on axis attention mechanism and related equipment

Abstract

A method for constructing an electromagnetic spectrum map based on an axis attention mechanism and related equipment are provided to solve the problems of large amount of detection data required by the electromagnetic spectrum map construction method and low construction precision. The method comprises: obtaining sparse electromagnetic data of a space to be reconstructed; obtaining an electromagnetic spectrum map of a wireless electromagnetic environment of the space to be reconstructed according to the sparse electromagnetic data and an electromagnetic data reconstruction network model based on the axis attention mechanism; wherein obtaining the electromagnetic spectrum map of the wireless electromagnetic environment of the space to be reconstructed according to the sparse electromagnetic data and the electromagnetic data reconstruction network model based on the axis attention mechanism comprises: extracting global features from the sparse electromagnetic data according to the sparse electromagnetic data and a global feature extraction sub-model; extracting local features from the sparse electromagnetic data according to the sparse electromagnetic data and a local feature extraction sub-model; and obtaining the wireless electromagnetic environment spectrum map of the space to be reconstructed according to the global features, the local features and a feature fusion module.

Description

Technical Field

[0001] One or more embodiments of this application relate to the field of geographic information technology, and in particular to an electromagnetic spectrum map construction method and related equipment based on axis attention mechanism. Background Technology

[0002] The electromagnetic spectrum is a continuous range describing electromagnetic waves of different frequencies and wavelengths, including radio waves, microwaves, infrared radiation, visible light, ultraviolet light, X-rays, and gamma rays. The electromagnetic spectrum has wide applications in communication, detection, and geographic information systems. In recent years, electromagnetic spectrum mapping technology has developed rapidly. By providing indicators such as received signal strength and power spectral density for each location in a geographic area, it has many applications, such as UAV communication, spectrum management, interference control, resource allocation, and network planning.

[0003] Electromagnetic spectrum maps are mainly composed of electromagnetic data collected by wireless sensors and mobile terminal devices distributed in space. However, due to limitations in physical environment and human resources, it is impossible to obtain electromagnetic data for every location point in the environment, thus making it impossible to construct a high-precision electromagnetic spectrum map.

[0004] To address this issue, related technologies have largely focused on building network models with different structures, such as Convolutional Neural Networks (CNNs), Residual Neural Networks (ResNets), and Transformers. While these technologies have improved reconstruction accuracy, they haven't considered the performance of models on small datasets. Generally, the larger the dataset, the more data features the model learns, resulting in smaller reconstruction errors. However, this also leads to increased computational resource consumption and a larger amount of data to be detected. Summary of the Invention

[0005] In view of this, the purpose of one or more embodiments of this application is to propose an electromagnetic spectrum map construction method and related device based on axis attention mechanism to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, one or more embodiments of this application provide a method for constructing an electromagnetic spectrum map based on an axis attention mechanism, including:

[0007] Acquire sparse electromagnetic data of the space to be reconstructed;

[0008] The electromagnetic spectrum map of the radio electromagnetic environment of the space to be reconstructed is obtained based on the sparse electromagnetic data and the electromagnetic data reconstruction network model based on the axis attention mechanism; the electromagnetic data reconstruction network model includes a global feature extraction sub-model, a local feature extraction sub-model, and a feature fusion module;

[0009] The process of obtaining a spectral map of the radio electromagnetic environment of the space to be reconstructed based on the sparse electromagnetic data and the electromagnetic data reconstruction network model using the axis attention mechanism includes:

[0010] Based on the sparse electromagnetic data and the global feature extraction sub-model, global features are extracted from the sparse electromagnetic data; the global features indicate the propagation characteristics of electromagnetic signals in long-distance space and reflect the overall distribution of electromagnetic data in the space to be reconstructed.

[0011] Based on the sparse electromagnetic data and the local feature extraction sub-model, local features are extracted from the sparse electromagnetic data; the local features indicate the propagation characteristics of electromagnetic signals in short-distance space and reflect the changes between electromagnetic data in the space to be reconstructed.

[0012] Based on the global features, the local features, and the feature fusion module, a radio electromagnetic environment spectrum map of the space to be reconstructed is obtained.

[0013] Optionally, the global feature extraction sub-model includes a first feature extraction layer built on a convolutional neural network, a second feature extraction layer built on an axial attention mechanism computation layer, and a first deconvolution layer; the axial attention mechanism computation layer is used to mine data patterns and data structures implicit in the sparse electromagnetic data that are related to the global features of the sparse electromagnetic data, so as to simulate the propagation law of electromagnetic signals and realize the reconstruction of the radio electromagnetic environment;

[0014] Based on the sparse electromagnetic data and the global feature extraction sub-model, global features are extracted from the sparse electromagnetic data, including:

[0015] The sparse electromagnetic data is input into the first feature extraction layer to obtain a feature map x of the sparse electromagnetic data based on convolution calculation. g The feature map x g This represents the basic global features extracted from the sparse electromagnetic data, wherein the basic global features include spatial propagation feature information and temporal propagation feature information contained in the sparse electromagnetic data; wherein, C g The feature map x represents g Number of feature channels, H g The feature map x represents g Height, W g The feature map x represents g The width;

[0016] The feature map x g Input the second feature extraction layer to extract features based on the feature map x. gGlobal attention weights for different feature channels in the feature map x g The global features of the sparse electromagnetic data are obtained by weighting; the global attention weights indicate the feature map x. g The importance of each feature channel;

[0017] The global features of the sparse electromagnetic matrix are input into the first deconvolution layer to obtain the global reconstructed feature map of the sparse electromagnetic matrix in a high dimension.

[0018] Optionally, the local feature extraction sub-model includes a third feature extraction layer and a second deconvolution layer constructed based on a convolutional neural network;

[0019] Based on the sparse electromagnetic data and the local feature extraction sub-model, local features are extracted from the sparse electromagnetic data, including:

[0020] The sparse electromagnetic data is processed by dividing it into blocks;

[0021] The sparse electromagnetic data after block processing is input into the third feature extraction layer to obtain local features of sparse electromagnetic data over short distances.

[0022] The local features of the sparse electromagnetic matrix are input into the second deconvolution layer to obtain the local reconstructed feature map of the sparse electromagnetic data in a high dimension.

[0023] Optionally, the feature fusion module includes a feature fusion layer and a data dimensionality reduction layer built based on a convolutional neural network;

[0024] Based on the global features, the local features, and the feature fusion module, a radio electromagnetic environment spectrum map of the space to be reconstructed is obtained, including:

[0025] The global reconstruction feature map and the local reconstruction feature map are input into the feature fusion layer to add the global reconstruction feature map and the local reconstruction feature map according to the feature channel number dimension to obtain high-dimensional electromagnetic environment spectrum map data.

[0026] The high-dimensional electromagnetic environment spectrum map data is input into the data dimensionality reduction layer to reduce the dimensionality of the electromagnetic environment spectrum map data, thereby obtaining the radio electromagnetic environment spectrum map of the space to be reconstructed.

[0027] Optionally, the global attention weights are determined based on relative position encoding, which indicates the feature map x. g The correlation between the distance between any elements and the relationship between those elements is used to capture the long-distance dependencies of the sparse electromagnetic data and to represent the spatial structure information of the sparse electromagnetic data.

[0028] Optionally, the global characteristics of the sparse electromagnetic data can be represented as:

[0029]

[0030] in, This represents the query vector, which is used to guide the focus area of ​​the second feature extraction layer. W Q This represents the query weight matrix, x represents the input data of the second feature extraction layer, and the input data is the basic global feature obtained by the first feature extraction layer based on the sparse electromagnetic data, k iw Let k represent the key vector, which is the query object of the query vector and is used to calculate similarity with the query vector. iw =W K x, W K Represents the key-value weight matrix, v iw Represents a value vector, v iw =W V x, W V This represents the query weight matrix, which is used to perform a weighted sum based on the similarity between the query vector and the key vector. and The relative position encoding parameters corresponding to the query vector, the key vector, and the value vector are used to guide the axis attention mechanism layer to focus on key information of non-local dependencies in the input data. Q S K S V1 and S V2 The perceptual parameter representing the degree of influence of the relative position encoding on the global feature learning process is used to control the influence of the learned relative position encoding on the global context encoding. i represents the height index of the input data, i∈{1,...,H}, j ​​represents the width index of the input data, j∈{1,...,W}, and w represents the index value that the second feature extraction layer needs to calculate.

[0031] Optionally, acquiring the sparse electromagnetic data of the space to be reconstructed includes:

[0032] Collect the power intensity values ​​of the electromagnetic signals in the space to be reconstructed;

[0033] The mask matrix of the space to be reconstructed is determined based on the power intensity value of the electromagnetic signal and the acquisition location of the electromagnetic signal. The mask matrix is ​​used to mark the sampling status of the power intensity value of the electromagnetic signal in the space to be reconstructed.

[0034] The power intensity value of the electromagnetic signal and the mask matrix are encapsulated into the sparse electromagnetic data.

[0035] Based on the same inventive concept, one or more embodiments of this application also provide an electromagnetic spectrum map construction device based on axis attention mechanism, including:

[0036] The data acquisition module is configured to acquire sparse electromagnetic data of the space to be reconstructed;

[0037] The radio electromagnetic environment reconstruction acquisition module is configured to obtain an electromagnetic spectrum map of the radio electromagnetic environment of the space to be reconstructed based on the sparse electromagnetic data and an electromagnetic data reconstruction network model based on the axis attention mechanism; the electromagnetic data reconstruction network model includes a global feature extraction sub-model, a local feature extraction sub-model, and a feature fusion module;

[0038] The process of obtaining a spectral map of the radio electromagnetic environment of the space to be reconstructed based on the sparse electromagnetic data and the electromagnetic data reconstruction network model using the axis attention mechanism includes:

[0039] Based on the sparse electromagnetic data and the global feature extraction sub-model, global features are extracted from the sparse electromagnetic data; the global features indicate the propagation characteristics of electromagnetic signals in long-distance space and reflect the overall distribution of electromagnetic data in the space to be reconstructed.

[0040] Based on the sparse electromagnetic data and the local feature extraction sub-model, local features are extracted from the sparse electromagnetic data; the local features indicate the propagation characteristics of electromagnetic signals in short-distance space and reflect the changes between electromagnetic data in the space to be reconstructed.

[0041] Based on the global features, the local features, and the feature fusion module, a radio electromagnetic environment spectrum map of the space to be reconstructed is obtained.

[0042] Based on the same inventive concept, one or more embodiments of this application also provide an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the electromagnetic spectrum map construction method based on the axis attention mechanism as described in any of the above.

[0043] Based on the same inventive concept, one or more embodiments of this application also provide a non-transitory computer-readable storage medium storing computer instructions for causing the computer to execute any of the above-described electromagnetic spectrum map construction methods based on axis attention mechanism.

[0044] As can be seen from the above, the electromagnetic spectrum map construction method based on axis attention mechanism provided in one or more embodiments of this application obtains sparse electromagnetic data of the space to be reconstructed; obtains an electromagnetic spectrum map of the radio electromagnetic environment of the space to be reconstructed based on the sparse electromagnetic data and an electromagnetic data reconstruction network model based on axis attention mechanism; the electromagnetic data reconstruction network model includes a global feature extraction sub-model, a local feature extraction sub-model, and a feature fusion module. Specifically, obtaining the electromagnetic spectrum map of the radio electromagnetic environment of the space to be reconstructed based on the sparse electromagnetic data and the electromagnetic data reconstruction network model based on axis attention mechanism includes: extracting global features from the sparse electromagnetic data based on the sparse electromagnetic data and the global feature extraction sub-model; the global features indicate the propagation characteristics of electromagnetic signals in long-distance space, reflecting the overall distribution of electromagnetic data in the space to be reconstructed; extracting local features from the sparse electromagnetic data based on the sparse electromagnetic data and the local feature extraction sub-model; the local features indicate the propagation characteristics of electromagnetic signals in short-distance space, reflecting the changes between electromagnetic data in the space to be reconstructed; and obtaining the radio electromagnetic environment spectrum map of the space to be reconstructed based on the global features, the local features, and the feature fusion module.

[0045] By constructing an electromagnetic data reconstruction network model that includes a global feature extraction sub-model, a local feature extraction sub-model, and a feature fusion module, the feature extraction capability of the model is improved. High-precision electromagnetic spectrum map construction can be achieved based on electromagnetic data with sparse features without increasing reconstruction error or reconstruction accuracy.

[0046] The electromagnetic spectrum map construction device, electronic device, and computer-readable storage medium provided in this application can all implement the steps of the electromagnetic spectrum map construction method based on the axis attention mechanism described above, and therefore also have the beneficial effects of the electromagnetic spectrum map construction method based on the axis attention mechanism described above. Attached Figure Description

[0047] To more clearly illustrate the technical solutions in one or more embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only one or more embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0048] Figure 1 This is a flowchart illustrating one or more embodiments of the electromagnetic spectrum map construction method based on the axis attention mechanism of this application;

[0049] Figure 2This is a flowchart illustrating one or more embodiments of the electromagnetic spectrum map construction method based on the axis attention mechanism of this application;

[0050] Figure 3 This is a flowchart illustrating one or more embodiments of the electromagnetic spectrum map construction method based on the axis attention mechanism of this application;

[0051] Figure 4 This is a schematic diagram of the electromagnetic spectrum map construction device based on the axis attention mechanism according to one or more embodiments of this application;

[0052] Figure 5 This is a schematic diagram of the feedforward calculation process of the axial attention meter layer in one or more embodiments of this application;

[0053] Figure 6 This is a schematic diagram illustrating the data reconstruction process of one or more embodiments of this application;

[0054] Figure 7 This is a schematic diagram illustrating the reconstruction accuracy experimental results of one or more embodiments of this application;

[0055] Figure 8 This is a schematic diagram of the experimental results of reconstructing the error distribution according to one or more embodiments of this application;

[0056] Figure 9 This is a schematic diagram of the experimental results of reconstructing the error distribution according to one or more embodiments of this application;

[0057] Figure 10 This is a schematic diagram of the hardware structure of an electronic device according to one or more embodiments of this application. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of this disclosure clearer, the following detailed description is provided in conjunction with specific embodiments and the accompanying drawings.

[0059] It should be noted that, unless otherwise defined, the technical or scientific terms used in one or more embodiments of this application should have the ordinary meaning understood by one of ordinary skill in the art to which this disclosure pertains. The terms "first," "second," and similar terms used in one or more embodiments of this application do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0060] Electromagnetic spectrum maps provide a comprehensive overview of key information such as the distribution of electromagnetic energy and spectrum resources in the electromagnetic environment through multi-dimensional quantitative descriptions, including time, frequency, space, and field strength, combined with geographic information for visualization.

[0061] Based on whether prior information such as radiation sources and propagation models is required, traditional electromagnetic spectrum map construction methods can generally be divided into three categories: direct construction methods, indirect construction methods, and hybrid construction methods. Direct construction methods mainly include the nearest neighbor method, natural neighbor method, inverse distance weighting method, and spline method. These methods do not require prior information about the electromagnetic environment, such as radiation sources, and can quickly construct a complete spatial electromagnetic situation distribution map based on known electromagnetic data, but they cannot accurately reflect the actual propagation laws of electromagnetic waves. Indirect construction methods require prior information such as radiation sources and propagation models, and mainly include forward modeling based on complete prior information, estimation methods based on transmitter positions, differential received signal strength methods, and signal-to-noise ratio-assisted methods. However, due to the existence of different types of obstacles and radiation sources with varying transmission frequencies and powers in the real environment, it is difficult to obtain all the prior information used for calculation. Typically, only some sparse electromagnetic data monitoring values ​​and information about some radiation sources can be obtained. Hybrid construction methods combine the advantages of the above two methods but do not solve the problem of difficulty in obtaining electromagnetic data.

[0062] However, the direct construction methods mentioned above, such as the nearest neighbor method, natural leader point method, and spline method, suffer from low construction accuracy and may lead to a reduction in the resolution of the electromagnetic spectrum map. When the distribution of electromagnetic data points is uneven, discontinuous abrupt changes will occur in the interpolation region, resulting in obvious step-like unevenness in the spectrum map. Furthermore, traditional methods are quite sensitive to measurement noise, and a single outlier can significantly impact the interpolation results. Electromagnetic spectrum map reconstruction techniques based on indirect construction methods also have a series of limitations. These methods rely on pre-set mathematical models, and the accuracy of these models is limited by the degree of understanding of the actual electromagnetic environment, which may lead to deviations between the reconstructed spectrum map and the real situation. Secondly, the estimation process of model parameters may be affected by the quality and quantity of measurement data, thus introducing estimation errors and affecting the accuracy of the map. Data sparsity means that with a limited number of measurement points, the reconstructed spectrum map may not fully reflect the distribution characteristics of the entire spectrum. Most importantly, these methods have insufficient generalization ability, meaning that they may perform well under specific conditions but poorly under different environments or conditions. In summary, neither of these methods can capture electromagnetic propagation characteristics in real time. They merely reconstruct data through geospatial interpolation algorithms or numerical calculations, rather than truly learning the propagation characteristics of electromagnetic waves in the real physical environment. Therefore, they have inherent limitations in constructing electromagnetic spectrum maps.

[0063] To address the shortcomings of the aforementioned methods, learn electromagnetic propagation characteristics as much as possible, and improve the accuracy of electromagnetic spectrum map reconstruction, a method for constructing electromagnetic spectrum using deep learning technology has been proposed.

[0064] In recent years, with the development of artificial intelligence technology, those skilled in the art have discovered that using machine learning and deep learning algorithms to construct electromagnetic spectrum maps can extract spatial features of electromagnetic propagation from a large amount of spectrum detection data, helping to improve the accuracy and precision of spectrum map construction. Therefore, this type of method has become a recent research hotspot. Its core idea is to use electromagnetic measurement datasets obtained under different environments to learn related electromagnetic propagation phenomena in physical spatial structures, such as shadows, reflections, and diffractions. In this way, the model can intuitively learn how these phenomena evolve in space, thereby significantly reducing the number of measurements required to achieve the desired accuracy.

[0065] Specifically, the related technologies propose a deep learning-based method for estimating urban outdoor environment spectrum maps. This method models the missing electromagnetic spectrum data completion problem as a tensor completion problem and solves the completion task using a deep neural network architecture with an autoencoder, achieving high construction accuracy. The related technologies also propose a 3D wireless map reconstruction scheme based on Generative Adversarial Networks (GANs). This model employs ResNet and dilated convolutional modules, utilizing data collected by UAVs to reconstruct a complete spectrum map. This model architecture has achieved good results in spectrum map reconstruction tasks in sensor networks and other different scenarios. Furthermore, the related technologies utilize a U-Net network model to learn the complex relationships between 3D objects, signal strength measurements, and the entire signal strength field.

[0066] However, the above methods still have some problems. The autoencoder structure was the first to apply deep learning networks to electromagnetic map reconstruction. Its training dataset comes from data generated by empirical path loss models in urban environments and data generated using ray tracing algorithms, comprehensively representing the real radio propagation environment. However, this method requires a large-scale training dataset, consuming significant hardware resources. Furthermore, since the autoencoder network uses fully convolutional layers as feature extraction layers, due to the inherent inductive bias of convolutional architectures, they cannot understand long-range dependencies in the input data. Convolutional kernels can only focus on local data features, causing the network to ignore global content. Generative Adversarial Networks (GANs) introduce two networks for adversarial training compared to previous algorithms: a generator generates data, and a discriminator evaluates the error between the generator's data and the original data, thereby continuously improving the generator's ability during the adversarial process. However, this type of method is not suitable for the practical situation of limited electromagnetic data detection in electromagnetic propagation environments. It usually requires a large number of samples to achieve good model performance, making training difficult.

[0067] In other words, this type of method belongs to data-driven algorithms. Unlike the construction techniques mentioned above, data-driven methods directly use raw electromagnetic data for feature learning and reconstruction prediction, completing the reconstruction task by mining the feature correlations between input data. As mentioned in the background, the quantity and density of the electromagnetic data collected are directly related to the accuracy of the final generated electromagnetic spectrum map. However, in practical applications, due to the distribution of buildings and obstacles, detection equipment cannot cover the entire detection area, resulting in only limited electromagnetic data being acquired. Therefore, how to utilize these limited and discrete electromagnetic data to construct a complete and continuous electromagnetic spectrum map has become the focus of current research.

[0068] Therefore, this application proposes an electromagnetic spectrum map construction method based on axis attention mechanism. While ensuring that the reconstruction error does not increase and the reconstruction accuracy does not decrease, by designing a new network with strong feature extraction capability and low complexity, the training cost can be effectively reduced and a better electromagnetic spectrum map reconstruction effect can be achieved.

[0069] refer to Figure 1 The electromagnetic spectrum map construction method based on axis attention mechanism according to one or more embodiments of this application includes the following steps:

[0070] Step S101: Obtain sparse electromagnetic data of the space to be reconstructed.

[0071] Step S102: Based on the above sparse electromagnetic data and the electromagnetic data reconstruction network model based on the axis attention mechanism, obtain the electromagnetic spectrum map of the radio electromagnetic environment of the space to be reconstructed; the above electromagnetic data reconstruction network model includes a global feature extraction sub-model, a local feature extraction sub-model and a feature fusion module.

[0072] Among them, such as Figure 2 As shown, based on the aforementioned sparse electromagnetic data and the electromagnetic data reconstruction network model based on the axis attention mechanism, a spectrum map of the radio electromagnetic environment of the space to be reconstructed is obtained, including:

[0073] Step S201: Based on the above sparse electromagnetic data and the above global feature extraction sub-model, extract global features from the above sparse electromagnetic data; the above global features indicate the propagation characteristics of electromagnetic signals in long-distance space and reflect the overall distribution of electromagnetic data in the space to be reconstructed.

[0074] Step S202: Based on the above sparse electromagnetic data and the above local feature extraction sub-model, extract local features from the above sparse electromagnetic data; the above local features indicate the propagation characteristics of electromagnetic signals in short-distance space and reflect the changes between electromagnetic data in the above space to be reconstructed.

[0075] Step S203: Based on the above global features, the above local features and the above feature fusion module, obtain the radio electromagnetic environment spectrum map of the space to be reconstructed.

[0076] Based on the above, in the technical solution of this application, if Figure 3 As shown, the data processing flow includes the following steps:

[0077] Step S301: Obtain sparse electromagnetic data of the space to be reconstructed;

[0078] Step S302: Extract global features and local features from the sparse electromagnetic data based on the global feature extraction sub-model and the local feature extraction sub-model;

[0079] Step S303: The global features and the local features are fused to obtain a high-dimensional radio electromagnetic environment spectrum map of the space to be reconstructed.

[0080] Step S304: Perform dimensionality reduction on the high-dimensional radio electromagnetic environment spectrum map to obtain the final radio electromagnetic environment spectrum map of the space to be reconstructed.

[0081] Specifically, in the embodiments of this application, the sparse electromagnetic data in step S301 may include: the power intensity value of the electromagnetic signal in the space to be reconstructed and the mask matrix corresponding to the power intensity value of the electromagnetic signal. The mask matrix can be a binary matrix, with elements taking values ​​of 0 or 1, where 0 indicates that the electromagnetic signal intensity value at that location is missing, and 1 indicates that the signal intensity value at that location is known and has been collected. The mask matrix is ​​used to mark the sampling status of the power intensity value of the electromagnetic signal in the space to be reconstructed. In the embodiments of this application, the sparse electromagnetic data may also include an information matrix, which may include a radiation source location information matrix, a propagation scene information marking matrix, etc., and is usually determined based on the information available in the actual application scenario. This application does not limit the data content of the information matrix. The mask matrix and the information matrix help the electromagnetic data reconstruction network model based on the axis attention mechanism to better identify the characteristics of the power intensity value of the electromagnetic signal, thereby improving the final reconstruction accuracy.

[0082] Based on the above, in the embodiments of this application, the acquisition of sparse electromagnetic data may include the following steps: first, acquiring the power intensity value of the electromagnetic signal in the space to be reconstructed; then, mapping the power intensity value of the electromagnetic signal to a pre-set matrix structure according to the position to obtain the power intensity value of the electromagnetic signal in matrix representation, and generating a mask matrix accordingly.

[0083] The matrix representation of electromagnetic data can include mapping the power intensity values ​​of sampled electromagnetic signals to a pre-defined matrix structure according to their spatial location. The rows and columns of the matrix correspond to specific locations within a spatial grid, and the matrix elements represent the power intensity values ​​of the electromagnetic signals within their respective grid cells. This step achieves the matrix representation of spatial data, laying the foundation for subsequent calculations and analysis.

[0084] Furthermore, since the model in this application is based on deep learning technology, and the basic unit for matrix operations in the electromagnetic data reconstruction network model based on axis attention mechanism is a tensor, the electromagnetic data can also be encapsulated as a tensor and then input into the model in the embodiments of this application.

[0085] The electromagnetic data reconstruction network model based on axis attention mechanism proposed in this application includes a global feature extraction sub-model, a local feature extraction sub-model, and a feature fusion module.

[0086] The global feature extraction sub-model extracts global features from sparse electromagnetic data. These global features indicate the propagation characteristics of electromagnetic signals over long distances, such as path loss and reflection. The local feature extraction sub-model extracts local features from sparse electromagnetic data. These local features indicate the propagation characteristics of electromagnetic signals over short distances, such as multipath effects and electromagnetic interference. Global features provide the model with background information about the data, while local features help the model capture micro-level changes in the data. Combining the two helps enhance the model's robustness, reduce the impact of a single feature, and improve inference capabilities. Furthermore, multi-level feature representations can lead to richer representations, which helps optimize feature selection in complex tasks, further improving the model's efficiency and accuracy.

[0087] Specifically, in the embodiments of this application, the global feature extraction sub-model in step S302 may include a first feature extraction layer constructed based on a convolutional neural network, a second feature extraction layer constructed based on an axial attention mechanism computation layer, and a first deconvolution layer. The aforementioned axial attention mechanism computation layer is used to mine the data patterns and data structures implicit in the aforementioned sparse electromagnetic data that are related to the global features of the aforementioned sparse electromagnetic data, so as to simulate the propagation law of electromagnetic signals and realize the reconstruction of the radio electromagnetic environment.

[0088] Axis attention mechanism is an attention mechanism introduced in deep learning to enhance the model's ability to represent features in multidimensional data (such as images, videos, or other multi-channel data). Its main idea is to adaptively focus on more important features by applying individual and joint attention weights to features on different axes (dimensions), thereby improving the model's ability to capture key information.

[0089] In implementing this application, the applicant discovered that, even with relatively small electromagnetic spectrum datasets, a deep learning algorithm network model based on axis attention mechanisms can effectively capture the correlation features between electromagnetic data, thereby enabling high-precision reconstruction of electromagnetic spectrum maps. Compared to previous research, this reduces the reliance on massive training data. Furthermore, the reduced number of network model parameters significantly saves hardware resources compared to previously proposed methods, allowing for rapid model training and derivation.

[0090] The design concept of the second feature extraction layer in this application is as follows:

[0091] The expression for the attention layer in the prior art can be:

[0092]

[0093] in k hw =WK x, v hw =W V x represents the query vector, key vector, and value vector, respectively, and v is the value vector. hw The global attention weights are calculated based on the softmax function, where x represents the input feature map data. H represents the height of the input data, and W represents the width of the input data. Here, the attention mechanism allows the model to weight different parts of the input data during processing, assigning different weights to different components. This allows the model to focus more on key information in the input sequence and reduce attention to secondary information. Therefore, unlike convolutional neural networks, self-attention mechanisms can capture non-local information from the entire feature map. However, calculating such attention weights is very time-consuming and consumes significant hardware resources.

[0094] To overcome the computational complexity of calculating attention weights, traditional self-attention can be decomposed into two self-attention modules. The first module performs self-attention operations on the height axis of the feature map, and the second module operates on the width axis. This computational approach is called axial attention. Therefore, axial attention applied on both the height and width axes can effectively simulate the original self-attention mechanism while offering higher computational efficiency, making it advantageous for processing electromagnetic data. Furthermore, positional encoding is applied to all query vectors, key vectors, and value vectors, enabling the capture of long-distance features between data points with precise positional information. The axial self-attention mechanism with positional encoding can be represented as:

[0095]

[0096] Considering that the aforementioned axis attention mechanism with positional encoding typically requires large-scale datasets for training, and given the inaccuracy of the learned electromagnetic propagation features and the relative positional encoding in the axis attention mechanism, this application proposes an improved axis attention calculation module. Specifically, it optimizes the training process of different positional vector encodings during electromagnetic data model training. This module learns the importance of positional encodings among different computational features and incorporates a perceptual parameter mechanism to control the importance of the data features represented by each relative positional encoding. This helps the network understand the long-distance relationships between data during training, calculate the correlation values ​​between different positional vectors, and ultimately select the most important data for the final output to complete the reconstruction and prediction task, thereby improving the model's expressive power and performance.

[0097] The improved axis self-attention mechanism with added perceptual parameters can be represented as:

[0098]

[0099] In the above formula, This represents the query vector, which is used to guide the focus of the second feature extraction layer. W Q Let x represent the query weight matrix, and let x represent the input data of the second feature extraction layer. The input data is the basic global feature obtained by the first feature extraction layer based on the sparse electromagnetic data. k iw This represents the key vector, which is the query object of the query vector and is used to calculate the similarity with the query vector. iw =W K x, W K Represents the key-value weight matrix, v iw Represents a value vector, v iw =W V x, W V This represents the query weight matrix, which is used to perform a weighted sum based on the similarity between the query vector and the key vector. and The relative position encoding parameters corresponding to the above query vector, key vector, and value vector are used to guide the above axis attention mechanism layer to focus on key information of non-local dependencies in the above input data. Q S K S V1 and S V2 The perceptual parameter representing the degree of influence of the above relative position encoding on the global feature learning process is used to control the influence of the learned relative position encoding on the global context encoding. i represents the height index of the input data, i∈{1,...,H}, j ​​represents the width index of the input data, j∈{1,...,W}, and w represents the index value that the second feature extraction layer needs to calculate.

[0100] It is important to note that in the embodiments of this application, if the electromagnetic data features represented by a relative position code are learned accurately, the sensing parameter mechanism will assign it a relatively high weight compared to codes that are not learned accurately. Here, inaccurate codes can be understood as regions in the electromagnetic data where the received signal strength values ​​are relatively low; this part of the data usually has the least impact on the final reconstructed data.

[0101] Therefore, in the embodiments of this application, the features of the global feature extraction sub-model are as follows: Figure 5As shown. The global feature extraction sub-model includes a first feature extraction layer based on a convolutional neural network, a second feature extraction layer based on an axial attention mechanism computation layer, and a first deconvolution layer. The axial attention mechanism computation layer is used to mine the data patterns and data structures implicit in the sparse electromagnetic data that are related to the global features of the sparse electromagnetic data, so as to simulate the propagation law of electromagnetic signals and realize the reconstruction of the radio electromagnetic environment.

[0102] Thus, in the embodiments of this application, extracting global features from the sparse electromagnetic data based on the above-mentioned sparse electromagnetic data and the above-mentioned global feature extraction sub-model includes:

[0103] The aforementioned sparse electromagnetic data is input into the first feature extraction layer to obtain a feature map x of the sparse electromagnetic data based on convolution calculation. g The above feature map x g This represents the basic global features extracted from the aforementioned sparse electromagnetic data. These basic global features include the spatial propagation and temporal propagation characteristics inherent in the sparse electromagnetic data; wherein, C g The above feature map x represents g Number of feature channels, H g Figure x above represents the above figure. g Height, W g The above feature map x represents g The width;

[0104] The above feature map x g Input the second feature extraction layer described above, to obtain the feature map x. g Global attention weights for different feature channels in the feature map x g The global features of the aforementioned sparse electromagnetic data are obtained through weighting; the aforementioned global attention weights indicate the aforementioned feature map x. g The importance of each feature channel;

[0105] The global features of the sparse electromagnetic matrix are input into the first deconvolution layer to obtain the global reconstructed feature map of the sparse electromagnetic matrix in a high dimension.

[0106] It should be noted that in the embodiments of this application, the global attention weights are determined based on relative position encoding, and the relative position encoding indicates the feature map x. g The correlation between the distance between any elements and the relationship between those elements is used to capture the long-distance dependencies of the sparse electromagnetic data and to represent the spatial structure information of the sparse electromagnetic data.

[0107] In the embodiments of this application, the local feature extraction sub-model in step S303 includes a third feature extraction layer and a second deconvolution layer constructed based on a convolutional neural network.

[0108] Thus, based on the aforementioned sparse electromagnetic data and the aforementioned local feature extraction sub-model, local features are extracted from the aforementioned sparse electromagnetic data, including:

[0109] The above sparse electromagnetic data is processed by dividing it into blocks;

[0110] The sparse electromagnetic data after block processing is input into the third feature extraction layer mentioned above to obtain the local features of sparse electromagnetic data over short distances.

[0111] The local features of the sparse electromagnetic matrix are input into the second deconvolution layer to obtain the local reconstructed feature map of the sparse electromagnetic data in a high dimension.

[0112] In the embodiments of this application, the feature fusion module of step S304 includes a feature fusion layer and a data dimensionality reduction layer constructed based on a convolutional neural network.

[0113] Thus, based on the aforementioned global features, local features, and feature fusion module, the radio electromagnetic environment spectrum map of the space to be reconstructed is obtained, including:

[0114] The above-mentioned global reconstruction feature map and the above-mentioned local reconstruction feature map are input into the above-mentioned feature fusion layer to add the above-mentioned global reconstruction feature map and the above-mentioned local reconstruction feature map according to the feature channel number dimension, so as to obtain high-dimensional electromagnetic environment spectrum map data.

[0115] The high-dimensional electromagnetic environment spectrum map data is input into the data dimensionality reduction layer to reduce the dimensionality of the electromagnetic environment spectrum map data, thereby obtaining the radio electromagnetic environment spectrum map of the space to be reconstructed.

[0116] Based on the above, the process of obtaining the electromagnetic spectrum map of the radio electromagnetic environment of the space to be reconstructed through an electromagnetic data reconstruction network model based on an axis attention mechanism in the embodiments of this application is as follows: Figure 6 As shown.

[0117] In other words, in the embodiments of this application, firstly, in the global feature extraction network branch, the electromagnetic data is directly input into the network at its original size for feature extraction. The features captured by the model focus more on the correlation between electromagnetic data at different locations globally, rather than just between adjacent electromagnetic data, because in space, the positional information between points that are close to each other has a high correlation, which would cause information redundancy. Then, in the local feature extraction network, the original data needs to be partitioned first, dividing the original input data into 4×4 sizes, and the partitioned data is sequentially fed into the network for feature extraction. The results are then fused with the global data features to form the final output. This allows for more refined feature capture by the model, and the improvement and innovation of the computation and feature fusion ideas at the method level are applied to the electromagnetic spectrum map reconstruction task.

[0118] To verify the effectiveness of the technical solution in this application, the applicant conducted algorithm simulation analysis experiments. The electromagnetic data used in the simulation came from the training dataset used in related technologies. Among the methods compared, the nearest neighbor method and Kriging interpolation are traditional interpolation algorithms based on spatial geographic information, which perform interpolation calculations based on the correlation of different geographic location data in space. These methods were widely used in early research. Multikernal is a reconstruction algorithm based on kernel functions, and Autoencoder is a deep learning-based algorithm used in related technologies. The main comparative indicators include root mean square error (RMSE), the cumulative distribution curve of reconstruction error, and the comparison between model parameters and computational complexity. This simulation used PyTorch version 2.3.0 for model building and training, Python version 3.10.14, CUDA version 12.4, an Nvidia RTX A2000 GPU with 24GB of video memory, and an Intel C622 CPU with 64GB of RAM.

[0119] To study the accuracy of the technical solution proposed in this application relative to the reconstruction accuracy of traditional algorithms, comparative simulations were conducted using traditional nearest neighbor (KNN), kriging interpolation, multikernal, and autoencoder algorithms from deep learning algorithms, respectively, against the algorithm proposed in this patent. Figure 7The vertical axis represents the RMSE error between the reconstructed value and the true value, and the horizontal axis represents the electromagnetic data at different sampling rates. It can be seen that the algorithm proposed in this patent, trained on a dataset of only 100,000 data points, achieves higher reconstruction accuracy at lower sampling rates than the autoencoder-based model in related technologies. It also outperforms four other traditional reconstruction methods. This preliminarily proves that the algorithm proposed in this patent has higher performance in electromagnetic data reconstruction accuracy. Furthermore, even with a small dataset, it can fully capture and learn data features, overcoming the shortcomings of existing research algorithms that require a large amount of data for training. This greatly reduces the cost of electromagnetic data acquisition and saves hardware training resources.

[0120] In addition to verifying the change in reconstruction error of the comparison methods as the number of sampling points changes, this patent also uses cumulative distribution curves to more intuitively statistically analyze the error distribution among various comparison methods. According to Figure 8 The results show that the RMSE error between the electromagnetic map generated by deep learning and the real map does not exceed 4dB, which is better than the traditional reconstruction method. In particular, it can achieve high reconstruction accuracy when the number of sampling points is relatively small.

[0121] To better reconstruct electromagnetic data, this application designed a global feature extraction branch and a local feature extraction branch during model construction. To verify the rationality of this design, corresponding simulation experiments were designed, including reconstruction using only the global feature extraction network, reconstruction using only the local feature network, and the network combining global and local features as used in this patent. Figure 9 The simulation results show that the reconstruction error of using only the local feature extraction network is the highest, and the reconstruction error of using only the global feature extraction network is also about 1dB higher on average than that of the combined network. This fully demonstrates the rationality of using the global feature extraction network, which can effectively capture the electromagnetic propagation characteristics in electromagnetic data, thereby enabling high-precision data reconstruction.

[0122] Furthermore, the applicant compared the model's parameter count and computational complexity by calculating two parameters: FLOPS and params. FLOPS stands for Floating-point Operations Per Second, which is the number of floating-point operations performed per second. It is commonly used to represent the number of floating-point operations performed per second and is a measure of a model's computational power. A higher FLOPS value usually means that the model can perform calculations faster or process larger amounts of data, but it also consumes more memory resources. Params in model descriptions usually refer to the number of parameters in the model. It represents the sum of all trainable parameters in the model. Params characterize the model's complexity. As shown in Table 1, the more parameters a model has, the more computational resources and data are required for training. The proposed method reduces the number of model parameters by 12% and FLOPs by 69.92% compared to related technologies, achieving high-precision reconstruction with lower model complexity.

[0123] Table 1

[0124]

[0125] In summary, the technical solution of this application applies a deep neural network based on axis attention mechanism to generate an electromagnetic spectrum map, which can significantly reduce the number of measurements required to achieve a given estimation accuracy. The basic principle of deep neural networks is to model complex data features and relationships through multi-layer nonlinear transformations. Each layer performs weighted and nonlinear transformations on the input data, progressively extracting and representing abstract features from the data. This hierarchical feature extraction enables neural networks to handle large-scale and high-dimensional data, such as images, speech, and natural language. Significant features of deep neural networks include highly parallel computing power, the ability to process large amounts of data, and automatic adjustment of model parameters during training to optimize performance. During computation, the electromagnetic data is first converted into the format of unit tensors for neural network computation. Then, through spatial discretization, the data of the study area is divided into discrete tables to simulate spatially discrete electromagnetic data. The design motivation of the algorithm studied in this patent is the observation that electromagnetic detection data contains electromagnetic propagation laws, such as reflection, diffraction, and path loss. These features are embedded in a high-dimensional space, and neural networks can learn these feature laws to complete the task of electromagnetic spectrum map reconstruction.

[0126] The technical solution of this application can fully extract features from data with a small sample size, thus completing the task of electromagnetic spectrum map reconstruction. In its design, the algorithm primarily reduces the number of computations during attention value calculation. Without losing global feature calculations, it retains the current feature and its corresponding vectors on the horizontal and vertical axes, significantly reducing the number of parameters the model needs to learn. Furthermore, it introduces learnable positional encoding parameters, enabling smoother and faster training convergence. From a feature fusion perspective, this invention designs a method to extract global and local features from the input data and fuse these features as the final output. This allows for a certain degree of learning of the relationships between long-distance and short-distance relationships between input data, improving the final reconstruction accuracy. Finally, in terms of application scenarios, the method proposed in this patent belongs to a general model computation architecture, thus applicable to model training in different scenarios and tasks, including image classification, image segmentation, language and text processing and generation. Depending on the different task objectives, by formatting the input and output data and fine-tuning the parameters using the model architecture, feature value calculation method, and feature fusion architecture designed in this invention, the desired goals can be achieved.

[0127] It is understandable that this method can be executed by any device, equipment, platform, or cluster of devices with computing and processing capabilities.

[0128] It should be noted that the methods of one or more embodiments of this application can be executed by a single device, such as a computer or server. The methods of this embodiment can also be applied in a distributed scenario, where multiple devices cooperate to complete the process. In such a distributed scenario, one of these devices may execute only one or more steps of the methods of one or more embodiments of this application, and the multiple devices will interact with each other to complete the method described.

[0129] It should be noted that the above description describes specific embodiments of this application. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps described in the claims may be performed in a different order than that shown in the embodiments and still achieve the desired results. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0130] Based on the same inventive concept, and corresponding to any of the methods in the above embodiments, this application also provides an electromagnetic spectrum map construction device based on an axis attention mechanism. For example... Figure 4 As shown, the above-mentioned device includes:

[0131] The data acquisition module 11 is configured to acquire sparse electromagnetic data of the space to be reconstructed;

[0132] The radio electromagnetic environment reconstruction acquisition module 12 is configured to obtain an electromagnetic spectrum map of the radio electromagnetic environment of the space to be reconstructed based on the above-mentioned sparse electromagnetic data and the electromagnetic data reconstruction network model based on the axis attention mechanism; the above-mentioned electromagnetic data reconstruction network model includes a global feature extraction sub-model, a local feature extraction sub-model and a feature fusion module;

[0133] Specifically, based on the aforementioned sparse electromagnetic data and the electromagnetic data reconstruction network model based on the axis attention mechanism, a spectrum map of the radio electromagnetic environment of the space to be reconstructed is obtained, including:

[0134] Based on the above sparse electromagnetic data and the above global feature extraction sub-model, global features are extracted from the above sparse electromagnetic data; the above global features indicate the propagation characteristics of electromagnetic signals in long-distance space and reflect the overall distribution of electromagnetic data in the space to be reconstructed.

[0135] Based on the above sparse electromagnetic data and the above local feature extraction sub-model, local features are extracted from the above sparse electromagnetic data; the above local features indicate the propagation characteristics of electromagnetic signals in short-distance space and reflect the changes between electromagnetic data in the above space to be reconstructed.

[0136] Based on the aforementioned global features, local features, and feature fusion module, a radio electromagnetic environment spectrum map of the space to be reconstructed is obtained.

[0137] For ease of description, the above apparatus is described in terms of its functions, divided into various modules. Of course, when implementing one or more embodiments of this application, the functions of each module can be implemented in one or more software and / or hardware.

[0138] The apparatus described above is used to implement the corresponding methods in the foregoing embodiments and has the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0139] Figure 10 This embodiment illustrates a more specific hardware structure of an electronic device, which may include a processor 1010, a memory 1020, an input / output interface 1030, a communication interface 1040, and a bus 1050. The processor 1010, memory 1020, input / output interface 1030, and communication interface 1040 are interconnected internally via the bus 1050.

[0140] The processor 1010 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application.

[0141] The memory 1020 can be implemented in the form of ROM (Read Only Memory), RAM (Random Access Memory), static storage device, dynamic storage device, etc. The memory 1020 can store the operating system and other applications. When the technical solutions provided in the embodiments of this application are implemented by software or firmware, the relevant program code is stored in the memory 1020 and is called and executed by the processor 1010.

[0142] The input / output interface 1030 is used to connect input / output modules to realize information input and output. Input / output modules can be configured as components within the device (not shown in the figure) or externally connected to the device to provide corresponding functions. Input devices may include keyboards, mice, touchscreens, microphones, various sensors, etc., while output devices may include displays, speakers, vibrators, indicator lights, etc.

[0143] The communication interface 1040 is used to connect a communication module (not shown in the figure) to enable communication between this device and other devices. The communication module can communicate via wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0144] Bus 1050 includes a pathway for transmitting information between various components of the device, such as processor 1010, memory 1020, input / output interface 1030, and communication interface 1040.

[0145] It should be noted that although the above-described device only shows the processor 1010, memory 1020, input / output interface 1030, communication interface 1040, and bus 1050, in specific implementations, the device may also include other components necessary for normal operation. Furthermore, those skilled in the art will understand that the above-described device may only include the components necessary for implementing the embodiments of this application, and not necessarily all the components shown in the figures.

[0146] The electronic devices described above are used to implement the corresponding methods in the foregoing embodiments and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0147] The computer-readable medium of this embodiment includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media 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 memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0148] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of this disclosure (including the claims) is limited to these examples; within the framework of this disclosure, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of one or more embodiments of this application as described above, which are not provided in detail for the sake of brevity.

[0149] Additionally, to simplify the description and discussion, and to avoid obscuring one or more embodiments of this application, the well-known power / ground connections to integrated circuit (IC) chips and other components may or may not be shown in the provided drawings. Furthermore, the apparatus may be illustrated in block diagram form to avoid obscuring one or more embodiments of this application, and this also takes into account the fact that the details of implementation of these block diagram apparatuses are highly dependent on the platform on which one or more embodiments of this application will be implemented (i.e., these details should be fully understood by those skilled in the art). While specific details (e.g., circuits) have been set forth to describe exemplary embodiments of this disclosure, it will be apparent to those skilled in the art that one or more embodiments of this application may be implemented without these specific details or with variations thereof. Therefore, these descriptions should be considered illustrative rather than restrictive.

[0150] Although this disclosure has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory architectures (e.g., dynamic RAM (DRAM)) may be used with the embodiments discussed.

[0151] One or more embodiments of this application are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of one or more embodiments of this application should be included within the protection scope of this disclosure.

Claims

1. A method for constructing an electromagnetic spectrum map based on an axis attention mechanism, characterized in that, The method comprises: acquiring sparse electromagnetic data of a space to be reconstructed; obtaining an electromagnetic spectrum map of a wireless electromagnetic environment of the space to be reconstructed based on an axis attention mechanism based electromagnetic data reconstruction network model according to the sparse electromagnetic data and the axis attention mechanism based electromagnetic data reconstruction network model; the electromagnetic data reconstruction network model comprises a global feature extraction sub-model, a local feature extraction sub-model and a feature fusion module; wherein obtaining the spectrum map of the wireless electromagnetic environment of the space to be reconstructed based on the sparse electromagnetic data and the axis attention mechanism based electromagnetic data reconstruction network model comprises: extracting global features from the sparse electromagnetic data according to the sparse electromagnetic data and the global feature extraction sub-model; the global features indicate the propagation characteristics of electromagnetic signals in a long-distance space and reflect the overall distribution of electromagnetic data in the space to be reconstructed; extracting local features from the sparse electromagnetic data according to the sparse electromagnetic data and the local feature extraction sub-model; the local features indicate the propagation characteristics of electromagnetic signals in a short-distance space and reflect the changes between electromagnetic data in the space to be reconstructed; obtaining the wireless electromagnetic environment spectrum map of the space to be reconstructed according to the global features, the local features and the feature fusion module; the global feature extraction sub-model comprises a first feature extraction layer constructed based on a convolutional neural network, a second feature extraction layer constructed based on an axis attention mechanism calculation layer and a first deconvolution layer; the axis attention mechanism calculation layer is used to mine the data patterns and data structures related to the global features of the sparse electromagnetic data and hidden in the sparse electromagnetic data, so as to simulate the propagation law of electromagnetic signals and realize the reconstruction of the wireless electromagnetic environment; extracting global features from the sparse electromagnetic data according to the sparse electromagnetic data and the global feature extraction sub-model comprises: input the sparse electromagnetic data into the first feature extraction layer to obtain a feature map of the sparse electromagnetic data based on convolution calculation , the feature map represents the extracted basic global features of the sparse electromagnetic data, and the basic global features contain spatial propagation feature information and time propagation feature information contained in the sparse electromagnetic data; wherein, , represents the number of feature channels of the feature map , represents the height of the feature map , represents the width of the feature map ; The feature map is obtained by performing feature extraction on the sparse electromagnetic data by using the first feature extraction layer. The second feature extraction layer is used to obtain a feature map according to the feature map The global attention weight of different feature channels in the feature map is used to weight the feature map The global feature of the sparse electromagnetic data is obtained; the global attention weight indicates the importance of each feature channel in the feature map The global feature of the sparse electromagnetic data is obtained; the global attention weight indicates the importance of each feature channel in the feature map inputting the global features of the sparse electromagnetic data into the first deconvolution layer to obtain the global reconstruction feature map of the sparse electromagnetic data in high dimension; The global attention weight is determined based on relative position encoding, the relative position encoding indicating the feature map The distance between any element and the relationship between the elements are associated to capture the long-distance dependence relationship of the sparse electromagnetic data and represent the spatial structure information of the sparse electromagnetic data. the global features of the sparse electromagnetic data are represented as: ； wherein, denotes a query vector, the query vector is used to guide the attention of the second feature extraction layer, , denotes a query weight matrix, denotes input data of the second feature extraction layer, the input data is a basic global feature obtained by the first feature extraction layer according to the sparse electromagnetic data, denotes a key vector, the key vector is a query object of the query vector, and is used for similarity calculation with the query vector, , denotes a key value weight matrix, denotes a value vector, , denotes a value query weight matrix, the query weight matrix is used for weighted summation according to the similarity between the query vector and the key vector, , and denotes the relative position encoding parameters corresponding to the query vector, the key vector and the value vector, which are used to guide the axis attention mechanism layer to pay attention to the key information of the non-local dependent relationship in the input data, , , and denotes the perception parameter of the influence degree of the relative position encoding on the global feature learning process, which is used to control the influence of the learned relative position encoding on the global context encoding, denotes a height index of input data, , denotes a width index of input data, , denotes an index value required to be calculated by the second feature extraction layer.

2. The method of claim 1, wherein, the local feature extraction sub-model comprises a third feature extraction layer constructed based on a convolutional neural network and a second deconvolution layer; extracting local features from the sparse electromagnetic data according to the sparse electromagnetic data and the local feature extraction sub-model comprises: performing block processing on the sparse electromagnetic data; inputting the block-processed sparse electromagnetic data into the third feature extraction layer to obtain the local features of the sparse electromagnetic data in a short distance; inputting the local features of the sparse electromagnetic data into the second deconvolution layer to obtain the local reconstruction feature map of the sparse electromagnetic data in high dimension.

3. The method of claim 2, wherein, the feature fusion module comprises a feature fusion layer and a data dimension reduction layer constructed based on a convolutional neural network; obtaining the wireless electromagnetic environment spectrum map of the space to be reconstructed according to the global features, the local features and the feature fusion module comprises: input the global reconstruction feature map and the local reconstruction feature map into the feature fusion layer to add the global reconstruction feature map and the local reconstruction feature map according to a feature channel number dimension, to obtain high-dimensional electromagnetic environment spectrum map data; input the high-dimensional electromagnetic environment spectrum map data into the data dimension reduction layer to reduce the dimension of the electromagnetic environment spectrum map data, to obtain a radio electromagnetic environment spectrum map of the space to be reconstructed.

4. The method of claim 1, wherein, The sparse electromagnetic data of the space to be reconstructed includes: collecting power intensity values of electromagnetic signals in the space to be reconstructed; determining a mask matrix of the space to be reconstructed according to the power intensity values of the electromagnetic signals and the collection positions of the electromagnetic signals, the mask matrix being used to mark the sampling situation of the power intensity values of the electromagnetic signals in the space to be reconstructed; encapsulating the power intensity values of the electromagnetic signals and the mask matrix into the sparse electromagnetic data.

5. An electromagnetic spectrum map construction device based on an axis attention mechanism, characterized by, It includes: a data acquisition module configured to acquire sparse electromagnetic data of a space to be reconstructed; a radio electromagnetic environment reconstruction acquisition module configured to obtain an electromagnetic spectrum map of a radio electromagnetic environment of the space to be reconstructed according to the sparse electromagnetic data and an electromagnetic data reconstruction network model based on an axis attention mechanism; the electromagnetic data reconstruction network model includes a global feature extraction sub-model, a local feature extraction sub-model, and a feature fusion module; wherein obtaining the spectrum map of the radio electromagnetic environment of the space to be reconstructed according to the sparse electromagnetic data and the electromagnetic data reconstruction network model based on the axis attention mechanism includes: extracting global features from the sparse electromagnetic data according to the sparse electromagnetic data and the global feature extraction sub-model; the global features indicate the propagation characteristics of electromagnetic signals in a long-distance space and reflect the overall distribution of electromagnetic data in the space to be reconstructed; extracting local features from the sparse electromagnetic data according to the sparse electromagnetic data and the local feature extraction sub-model; the local features indicate the propagation characteristics of electromagnetic signals in a short-distance space and reflect the changes between electromagnetic data in the space to be reconstructed; obtaining the radio electromagnetic environment spectrum map of the space to be reconstructed according to the global features, the local features, and the feature fusion module; the global feature extraction sub-model includes a first feature extraction layer constructed based on a convolutional neural network, a second feature extraction layer constructed based on an axis attention mechanism calculation layer, and a first deconvolution layer; the axis attention mechanism calculation layer is used to mine data patterns and data structures related to global features of the sparse electromagnetic data and implicitly contained in the sparse electromagnetic data, to simulate the propagation law of electromagnetic signals and realize the reconstruction of the radio electromagnetic environment; extracting global features from the sparse electromagnetic data according to the sparse electromagnetic data and the global feature extraction sub-model includes: input the sparse electromagnetic data into the first feature extraction layer to obtain a feature map of the sparse electromagnetic data based on convolution calculation , the feature map represents the extracted basic global features of the sparse electromagnetic data, and the basic global features contain spatial propagation feature information and time propagation feature information contained in the sparse electromagnetic data; wherein, , represents the number of feature channels of the feature map , represents the height of the feature map , represents the width of the feature map . extracting the feature map the second feature extraction layer, to obtain global features of the sparse electromagnetic data according to the feature map the global attention weights of different feature channels in the feature map weighting to obtain the global features of the sparse electromagnetic data; the global attention weights indicate the importance of each feature channel in the feature map the global attention weights indicate the importance of each feature channel in the feature map inputting the global features of the sparse electromagnetic data into the first deconvolution layer to obtain a global reconstruction feature map of the sparse electromagnetic data in a high dimension; The global attention weight is determined based on relative position encoding, the relative position encoding indicating distances between any elements in the feature map The association of the distance between any elements and the relationship between the elements is used to capture the long-distance dependence relationship of the sparse electromagnetic data and represent the spatial structure information of the sparse electromagnetic data. the global features of the sparse electromagnetic data are represented as: ； wherein, denotes a query vector, the query vector being used to guide the attention of the second feature extraction layer, , denotes a key weight matrix, denotes input data of the second feature extraction layer, the input data being basic global features obtained by the first feature extraction layer according to the sparse electromagnetic data, denotes a key vector, the key vector being a query object of the query vector, and being used to perform similarity calculation with the query vector, , denotes a key value weight matrix, denotes a value vector, , denotes a value query weight matrix, the query weight matrix being used to perform weighted summation according to the similarity between the query vector and the key vector, , and denote relative position encoding parameters corresponding to the query vector, the key vector and the value vector, and being used to guide the axis attention mechanism layer to pay attention to key information of non-local dependency in the input data, , , and denote perception parameters of the influence degree of the relative position encoding on the global feature learning process, and being used to control the influence of the learned relative position encoding on the global context encoding, denotes a height index of input data, , denotes a width index of input data, , denotes an index value required to be calculated by the second feature extraction layer.

6. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the program to implement the method of any one of claims 1 to 4.

7. A non-transitory computer-readable storage medium, comprising: The non-transitory computer readable storage medium stores computer instructions for causing the computer to perform the method of any one of claims 1 to 4.

Images