Radio environment map estimation method based on generative adversarial network

Abstract

The invention provides a radio environment map estimation method based on a generative adversarial network, which belongs to the technical field of wireless communication, and comprises the following steps: collecting radio map original data at different sampling rates; filling the sparse data to the same size as the original map; inputting the filled radio map into a lightweight depth high-resolution module for image feature enhancement; the radio map after feature enhancement is input into a generator module, and feature information is further extracted in combination with a dual-attention multi-scale fusion module and Transform; inputting the extra condition information and the real or generated radio map into a discriminator to judge whether the radio map is true or false; designing a loss function, and updating the generator and the discriminator; and repeating until the loss value tends to be stable. According to the method, the reconstruction quality of the radio signal can be improved, the adaptability to complex building shielding and signal attenuation conditions is enhanced, the uplink and downlink communication overhead in the training process can be remarkably saved, the dependence on a large amount of prior information is reduced, and the method has a wide application prospect.

Description

Technical Field

[0001] The present invention belongs to the field of wireless network communications, and more particularly, relates to a radio environment map estimation method based on a generative adversarial network. Background Art

[0002] A radio map (RM) is a key data format used to depict the propagation status of wireless signals in geographic space, focusing on spatial dimensions and integrating multi-dimensional information such as time, frequency, and power. The primary function of a radio map is to present electromagnetic characteristics such as wireless signal strength, path loss, interference distribution, and power spectral density (PSD). It is a crucial tool for intelligent perception, resource scheduling, and interference management in wireless communication networks.

[0003] With the rapid development of emerging scenarios such as 6G communications, smart cities, and autonomous driving, real-time understanding of the radio environment has become fundamental. High-precision radio maps not only provide a data basis for base station deployment but also offer technical support for dynamic allocation of spectrum resources, monitoring and avoidance of wireless interference, and high-precision positioning of terminals. However, due to factors such as high sampling costs, sparse sensor deployment, and the dynamic changes in complex propagation environments, traditional methods struggle to achieve comprehensive modeling at large scales and fine granularity.

[0004] In existing technologies, some interpolation methods (such as Kriging and Radial Basis Function (RBF)) attempt to use existing observation data to complete unknown areas. However, these methods perform poorly in areas with building occlusion or drastic changes in path loss, and the estimation error is large. In recent years, deep learning technology has been introduced into the Radio Map Estimation (RME) task, such as autoencoders, U-Net structures, and Generative Adversarial Networks (GANs), which have demonstrated good data fitting capabilities. However, they generally face two major problems: one is excessive reliance on prior information, such as the location of the transmitter or the distribution of buildings, which is difficult to obtain in actual deployments; the other is the lack of the ability to model long-distance dependencies under sparse observations, making it difficult to balance local details with global consistency.

[0005] To address these issues, some research has begun introducing Transformer architectures to model global features or using generative models to improve the restoration of spectrograms. However, a unified architecture has yet to emerge that can simultaneously address the challenges of sparse sampling, building occlusions, and mapping in unknown environments. In particular, in real-world applications, building structure information is often missing or incomplete, placing higher demands on the model's adaptability and generalization capabilities.

[0006] Therefore, a radio environment map estimation method based on generative adversarial networks is proposed. The present invention aims to solve the problem of high-precision map construction in sparse observation and complex propagation environments through an innovative deep learning architecture, avoid dependence on prior information such as the location of the transmitter, and possess strong generalization ability and robustness, thereby improving the perception accuracy and reliability of the radio environment. Summary of the Invention

[0007] To address the above-mentioned problems in the prior art, the present invention provides a method for estimating a radio environment map based on a generative adversarial network. The technical problem to be solved by the present invention is achieved through the following technical solutions:

[0008] In a first aspect, the present invention provides a radio environment map estimation method based on a generative adversarial network, comprising:

[0009] S1: Collect the current radio map raw data at different sampling rates.

[0010] S2: Pad the sparse data to the same size as the original radio map.

[0011] S3: The sparsely sampled radio map is input into the designed lightweight DHR for image feature enhancement.

[0012] S4: The image feature enhanced data is input into the improved generator module, and DAMF and Transformer are combined to further extract feature information.

[0013] S5: The additional conditional information and the real or generated radio map are input into the discriminator for true or false judgment.

[0014] S6: Design a loss function to minimize the loss and update the generator and discriminator.

[0015] S7: Repeat S1 to S6 until the loss value stabilizes.

[0016] In a possible implementation, the specific step of S1 is: dividing the target area into N×N grid units, with each K iThe small area contains a sensor. Each sensor receives the radio signal data of the area. These data are recorded in the form of power spectrum density (PSD) of the signal. Specifically, the Kth i The sparse samples of the i-th region accepted by the sensor are (x i , p i ),in is the PSD of non-negative observations, is the corresponding two-dimensional coordinate. In addition, region i also has a corresponding city map The city map shows the building distribution information, where a value of 1 indicates that it is blocked by buildings. The sampling rate is set according to the actual scene, K i ≤ N × N. Generally, a lower sampling rate can simulate a low density of measurement points, while a higher sampling rate is used for more accurate area coverage and signal capture.

[0017] In one possible implementation, step S2 specifically includes: first, obtaining a sparsely observed radio signal dataset. These datasets are typically collected by sensors located in specific locations within a given area. The number of radio signal samples obtained is limited, and the coverage area of ​​these samples is much smaller than the entire target radio map area.

[0018] In order to enable these sparse data to be input into the subsequent estimation module, they must be padded to the same size as the original radio map. Specifically, the sparse observations (x i , p i ) is zero-filled to fill the observation area to the same size as the original radio map, i.e. The interpolation result will not affect other features. Therefore, when the input feature has m distribution information, the feature is encoded into two channels, namely z i ={x′ i , m i}, when there is no m distribution information, the feature is encoded as a channel, namely z i ={x′ i}, each channel size is N×N.

[0019] In one possible implementation, step S3 specifically involves inputting the zero-padded, sparsely sampled radio map data into a designed lightweight DHR module. The DHR module extracts fine-grained features from the low-resolution radio map, particularly high-frequency details, thereby improving radio map restoration quality, particularly under sparse observation conditions. The DHR module employs a five-layer shallow convolutional architecture, significantly reducing computational complexity and memory usage compared to traditional 20-layer super-resolution networks. Specifically, in addition to the input and output layers, each DHR layer contains 64 convolutional kernels of size 3×3×64. Batch normalization and ReLU activation functions are used to improve training efficiency. These designs optimize the model training process, reducing network training time and enhancing feature learning capabilities. In particular, the layer-by-layer residual connection mechanism preserves original feature information, avoids vanishing gradients, and promotes efficient propagation of high-frequency information, ensuring that the network can extract richer details from low-resolution input data.

[0020] In one possible implementation, step S4 specifically includes: inputting the processed data into an improved generator module, which adopts an encoder-decoder structure. The decoder gradually restores the resolution of the radio map through skip connections and five layers of upsampling convolution blocks. The encoder part of the generator maps the input data to a high-dimensional feature space through convolution blocks, and gradually reduces the spatial size through downsampling operations, while increasing the number of feature channels. To enhance the ability to extract signal features in complex environments, the encoder integrates DAMF. The DAMF module combines void convolution and attention mechanisms to extract multi-scale features with different receptive fields, enhancing the ability to model local, regional, and global changes in the signal. The DAMF module further introduces the PAM and CAM mechanisms, which can dynamically adjust the weights of the spatial and channel dimensions of the features, thereby enhancing the response of key areas and important channels and effectively suppressing noise interference.

[0021] The feature maps processed by the DAMF module are input into the Transformer module. The Transformer module uses a self-attention mechanism to model long-range dependencies, addressing the limitations of traditional CNNs in modeling global features. The Transformer employs a multi-head self-attention mechanism to effectively integrate local information from sparse observation points with global information from areas obscured by buildings. Combined with positional encoding, the Transformer maintains the spatial consistency of the radio map.

[0022] In one possible implementation, step S5 specifically involves inputting the generated radio map and additional conditional information into a discriminator for authenticity determination. The discriminator utilizes a lightweight, stable, and efficient four-layer convolutional network architecture, effectively reducing computational complexity and optimizing the stability of generator training. The discriminator distinguishes between radio maps generated by the generator and those derived from real measurement data by performing authenticity determination on the input data.

[0023] To enhance the discriminator's discriminative capabilities, a three-channel input mechanism was designed, consisting of a real or generated radio map and two channels of one-hot encoded prior information. The one-hot encoding provides regional geometric features and signal marker information, which helps the discriminator accurately determine the authenticity of input data in complex environments. Specifically, the input data is processed by a four-layer convolutional network, with each layer performing further feature extraction on the input.

[0024] The discriminator ultimately outputs a one-dimensional probability value, which determines whether the input radio map is real data or fake data generated by the generator. This output value is a number between 0 and 1. Values ​​closer to 1 indicate a greater probability that the data is real data, while values ​​closer to 0 indicate a greater probability that the data is generated data.

[0025] In a possible implementation, step S6 is specifically as follows: we use the following objective function to train the generator and the discriminator:

[0026]

[0027] where p data (r, I) is the joint distribution together with the prior knowledge, Represents the unique hot encoding in the two channels. To optimize the training model, the maximum-minimum joint optimization problem is split into two sub-problems, and the generator G and the discriminator D are trained iteratively. Specifically, the loss function of the discriminator is:

[0028] L D = -V(D,G) (2)

[0029] The loss function used to optimize the generator is:

[0030]

[0031] To help the improved generator learn a feature distribution closer to the real radio map, we redesigned the generator's loss function. The generator loss combines global reconstruction loss, total variation loss (TV), gradient similarity loss (GS), and building-constrained loss (BC) to achieve efficient generator learning.

[0032] The global reconstruction loss aims to ensure the overall consistency between the generated radio map and the real radio map by minimizing the mean squared error (MSE) between the generated map and the target map. Its formula is:

[0033]

[0034] Among them, r represents the real spectrum graph, Indicates the generation of a spectrum graph. is the Frobenius norm.

[0035] TV is used to smoothly generate images and is defined as follows:

[0036]

[0037] in, Indicates the generation of a spectrum graph pixel area.

[0038] The gradient information of the image can effectively capture the edge and texture information, so we introduce GS to ensure that the gradient distribution of the generated spectrum image is similar to the real image.

[21] To measure the similarity between gradients, the gradient similarity formula is as follows:

[0039]

[0040] Where CS(·) represents the cosine similarity. Represents the generated map gradient.

[0041] To enhance the structural perception and recovery capabilities of building areas in different scenarios, we propose a structure-preserving loss function L BC , which are applicable to the two proposed scenarios respectively. In the scenario where the building distribution is known, the building mask M is introduced. b ∈{0, 1} 256×256 , where the building area is 1 and the others are 0. When the building distribution information is unknown, we use the area with extremely low grayscale value in the predicted value as the pseudo occlusion area to construct a pseudo mask:

[0042]

[0043] Where ∈ is the grayscale threshold set, which is used to approximately represent the area where occlusion may occur.

[0044] We hope that the model will not produce false signals in the building area, the formula is:

[0045]

[0046] The final generator loss function is defined as:

[0047] L D =λ M L M +λ TV L TV +λ GS L GS +λ BC L BC (9)

[0048] In a second aspect, the present invention proposes a radio map estimation device based on a multi-stage collaborative training strategy, comprising:

[0049] The data acquisition module is used to collect the current radio map raw data according to different sampling rates.

[0050] The data padding module is used to pad the sparse data to the same size as the original radio map.

[0051] Feature enhancement module, which inputs the sparsely sampled radio map into the designed lightweight DHR for image feature enhancement.

[0052] The generator module is used to input the processed data into the improved generator module, combining DAMF and Transformer.

[0053] The discriminator module inputs the additional conditional information and the real or generated radio map into the discriminator for true or false judgment.

[0054] The training optimization module is used to design a loss function that minimizes the loss and updates the generator and discriminator until the loss value stabilizes and the training process is completed.

[0055] In a third aspect, the present invention also provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the method for estimating a radio environment map based on a generative adversarial network as described above is implemented.

[0056] In a fourth aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the above-described methods for estimating a radio environment map based on a generative adversarial network. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the examples or descriptions of the prior art. Obviously, the drawings described below are some examples of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0058] Figure 1 A realistic scenario for radio map estimation under building coverage;

[0059] Figure 2 A flowchart of a radio environment map estimation method based on a generative adversarial network provided by the present invention;

[0060] Figure 3 A schematic diagram of the structure based on the generative adversarial model;

[0061] Figure 4 1 is a schematic structural diagram of a radio environment map estimation method based on a generative adversarial network provided by the present invention;

[0062] Figure 5 It is a structural schematic diagram of the electronic equipment provided by the present invention. DETAILED DESCRIPTION

[0063] The following describes the embodiments of the present invention through specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention based on the contents of this specification. The present invention can also be implemented or applied through different specific embodiments, and the details in this specification can be modified and improved according to different perspectives and application scenarios without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following examples are only schematic illustrations to illustrate the basic concept of the present invention. In the absence of conflict, the examples and the features therein can be combined with each other.

[0064] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual product images. Therefore, they should not be construed as limiting the present invention. To better illustrate the embodiments of the present invention, certain components in the accompanying drawings may be omitted, enlarged, or reduced in size, and do not represent the actual dimensions of the products. Certain structures well known to those skilled in the art and their descriptions may be omitted from the accompanying drawings. Such omissions are reasonable and understandable.

[0065] It should be noted that if Figure 1 As shown, the scenario considered by the present invention is the radio map estimation problem based on real building coverage. In this scenario, multiple radio towers, mobile phones and other communication equipment are distributed in an urban area, and sensors are used to collect radio signals in the area. The propagation of these signals is affected by factors such as building occlusion, reflection, and diffraction, resulting in complex and uneven signal distribution. In addition, due to the presence of building occlusion in the environment, the propagation path and intensity of the signal will vary greatly in different areas, so it is necessary to combine complex modeling algorithms for signal estimation. The radio map estimation method in this scenario aims to use these sparse observation data to complement the spatial distribution of wireless signals and solve the uncertainty caused by occlusion and propagation loss. Through real-time data updating and storage, it can provide accurate data support for wireless communication network optimization, spectrum management, interference detection and wireless positioning.

[0066] Figure 2 The flowchart of the radio environment map estimation method based on the generative adversarial network provided by the present invention is as follows: Figure 2 As shown, the method is:

[0067] S201: Collecting current radio map raw data according to different sampling rates.

[0068] It should be noted that in order to adapt to the scene efficiently, the target area is divided into N×N grid units, and each K i The small area contains a sensor. Each sensor receives the radio signal data of the area. These data are recorded in the form of power spectrum density (PSD) of the signal. Specifically, the sparse samples received by each sensor in the i-th area are (x i ,p i ),in is the PSD of non-negative observations, Corresponding two-dimensional coordinates. In addition, region i also has a corresponding city map The city map shows the building distribution information, where a value of 1 indicates that it is blocked by buildings. The sampling rate is set according to the actual scene, K i ≤ N × N. Generally, a lower sampling rate can simulate a low density of measurement points, while a higher sampling rate is used for more accurate area coverage and signal capture.

[0069] S202: Fill the sparse data to the same size as the original radio map.

[0070] It should be noted that in order to enable these sparse data to be input into the subsequent estimation module, they must be padded to the same size as the original radio map. Specifically, the sparse observation values ​​(x i , pi ) is zero-filled to fill the observation area to the same size as the original radio map, i.e. The interpolation result will not affect other features. Therefore, when the input feature has m distribution information, the feature is encoded into two channels, namely z i ={x′ i , m i}, when there is no m distribution information, the feature is encoded as a channel, namely z i ={x′ i}, each channel size is N×N.

[0071] S203: Input the sparsely sampled radio map into the designed lightweight DHR for image feature enhancement.

[0072] It should be noted that in order to improve the quality of the radio map, especially the signal recovery effect under sparse observation conditions. The DHR module adopts a 5-layer shallow convolution structure, which significantly reduces the computational complexity and memory usage compared to the traditional 20-layer super-resolution network. Specifically, in addition to the input layer and the output layer, each of the three layers of the DHR module contains 64 convolution kernels of size 3×3, combined with batch normalization and ReLU activation functions to improve training efficiency. These designs optimize the training process of the model, reduce the training time of the network and increase the feature learning ability. In particular, the layer-by-layer residual connection mechanism can retain the original feature information, avoid the gradient disappearance phenomenon, and promote the effective propagation of high-frequency information, thereby ensuring that the network can extract richer detail information from low-resolution input data.

[0073] S204: The image feature enhanced data is input into the improved generator module, and further feature information is extracted by combining DAMF and Transformer.

[0074] It should be noted that the generator module adopts an encoder-decoder structure. The decoder gradually restores the resolution of the radio map through skip connections and five layers of upsampling convolution blocks. The encoder part of the generator maps the input data to a high-dimensional feature space through convolution blocks, and gradually reduces the spatial size through downsampling operations, while increasing the number of feature channels. To enhance the ability to extract signal features in complex environments, the encoder integrates DAMF. The DAMF module combines void convolution and attention mechanisms to extract multi-scale features with different receptive fields, enhancing the ability to model local, regional, and global changes in the signal. The DAMF module further introduces the PAM and CAM mechanisms, which can dynamically adjust the weights of the spatial and channel dimensions of the features, thereby enhancing the response of key areas and important channels and effectively suppressing noise interference.

[0075] The feature maps processed by the DAMF module are input into the Transformer module. The Transformer module uses a self-attention mechanism to model long-range dependencies, addressing the limitations of traditional CNNs in modeling global features. The Transformer employs a multi-head self-attention mechanism to effectively integrate local information from sparse observation points with global information from areas obscured by buildings. Combined with positional encoding, the Transformer maintains the spatial consistency of the radio map.

[0076] S205: The additional condition information and the real or generated radio map are input into a discriminator for true or false judgment.

[0077] It should be noted that to enhance the discriminator's discriminative capabilities, a three-channel input mechanism was designed, consisting of a real or generated radio map and two channels of one-hot encoded prior information. The one-hot encoding provides regional geometric features and signal marker information, which helps the discriminator accurately determine the authenticity of input data in complex environments. Specifically, the input data is processed by a four-layer convolutional network, with each layer performing further feature extraction on the input.

[0078] The discriminator ultimately outputs a one-dimensional probability value, which determines whether the input radio map is real data or fake data generated by the generator. This output value is a number between 0 and 1. Values ​​closer to 1 indicate a greater probability that the data is real data, while values ​​closer to 0 indicate a greater probability that the data is generated data.

[0079] S206: Design a loss function to minimize the loss and update the generator and discriminator.

[0080] It should be noted that to encourage the improved generator to learn a feature distribution closer to the real radio map, we redesigned the generator's loss function. The generator loss combines global reconstruction loss, total variation loss (TV), gradient similarity loss (GS), and building-constrained loss (BC) to achieve efficient generator learning.

[0081] The final generator loss function is defined as:

[0082] L D =λ M L M +λ TV L TV ++λ GS L GS +λ BC L BC(10)

[0083] S207: Repeat S201 to S206 until the loss value stabilizes.

[0084] It should be noted that steps S201 to S206 are repeated until the loss values ​​of the generator and the discriminator tend to be stable and the training process is completed.

[0085] Figure 3 Schematic diagram of the radio map estimation structure based on the generative adversarial model.

[0086] It should be noted that the architecture of the generative adversarial model consists of two parts: a generator and a discriminator. The generator recovers a high-resolution map from sparsely sampled radio map data, using a Depth-Reduction (DHR) module, DAMF, and Transformer to enhance feature extraction and long-range signal dependency modeling. The discriminator uses a four-layer convolutional neural network to determine whether the input radio map is realistic. The generator and discriminator optimize radio map estimates through adversarial training, ultimately achieving high-precision radio map construction.

[0087] The following describes a radio map estimation device based on a generative adversarial model provided by the present invention. The radio map estimation device based on a generative adversarial model described below and the radio environment map estimation method based on a generative adversarial network described above can refer to each other.

[0088] Figure 4 4 is a schematic diagram of the structure of a radio estimation device based on a generative adversarial model provided by an embodiment of the present invention, comprising: a data acquisition module 401, a data filling module 402, a feature enhancement module 403, a generator module 404, a discriminator module 405, and a training optimization module 406, wherein:

[0089] The data acquisition module 401 is used to collect the current radio map raw data at different sampling rates; the data padding module 402 is used to pad the sparse data to the same size as the original radio map; the feature enhancement module 403 inputs the sparsely sampled radio map into the designed lightweight DHR for image feature enhancement; the generator module 404 inputs the processed data into the improved generator module, combining the DAMF and Transformer; the discriminator module 405 inputs additional conditional information and the real or generated radio map into the discriminator for true or false judgment; and the training optimization module 406 is used to design a loss minimization function and update the generator and discriminator until the loss value tends to stabilize, and the training process is completed.

[0090] Figure 5 An example of a hardware structure diagram of an electronic device is shown below. Figure 5As shown, the electronic device includes: a processor 501, a communication interface 502, a memory 503, and a communication bus 504. The modules exchange data via the communication bus. The processor 501 is configured to execute program instructions pre-stored in the memory 503 to implement a radio environment map estimation method based on a generative adversarial network. The method includes: collecting raw data of the current radio map at different sampling rates, padding the sparse data to the same size as the original radio map, and inputting the sparsely sampled radio map into a designed lightweight DHR for image feature enhancement. The processed data is input into an improved generator module, which combines DAMF and Transformer. The additional conditional information and the real or generated radio map are input into a discriminator for true or false judgment. A loss function is designed to minimize the loss, the generator and discriminator are updated, and S201 to S206 are repeated until the loss value stabilizes.

[0091] Furthermore, the memory 503 can be used to store program code, network model parameters, and intermediate results. The software instructions of the above functional units can be encapsulated as a computer program and deployed on various storage media. Such computer-readable storage media include, but are not limited to, USB flash drives, hard drives, ROM, RAM, magnetic tapes, and optical disks, and can be used to enable general-purpose computing devices (such as PCs, servers, and edge devices) to execute the above method steps and achieve automatic estimation and output of radio maps.

[0092] The present application example also provides a non-transitory computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the radio environment map estimation method based on a generative adversarial network as provided in the above methods is implemented.

[0093] The functional modules described in the above device examples are only schematic configurations, and their specific division should not be understood as a limitation on the physical structure. The units mentioned therein can be independent entity modules, or they can be integrated into the same hardware or software module. The various functional parts can be centrally deployed in the same device, or they can be distributed in multiple devices or device nodes as needed according to the actual application scenario. In actual deployment, all or part of the modules can be selected to be implemented according to functional requirements to achieve the technical effects of the present invention. Those skilled in the art can understand and apply the scheme of the present invention based on this without the need for creative work.

[0094] Through the description of the above embodiments, those skilled in the art will understand that each embodiment can be implemented through software combined with a general hardware platform, or can be completed using pure hardware. On this basis, the technical solution of the present invention can essentially be embodied in the form of a software product, stored in a computer-readable storage medium such as ROM, RAM, disk, CD, etc. This software product contains a number of instructions that can be used to enable a computing device (such as a personal computer, server, or network device) to execute the method described in the embodiments of the present invention or part of the steps thereof.

[0095] Finally, it should be noted that the above examples are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the above examples, those skilled in the art should understand that they can still modify the technical solutions described in the above examples or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the examples of the present invention.

Claims

1. A radio environment map estimation method based on a generative adversarial network, characterized in that: The steps include: S1: Collect the current radio map raw data at different sampling rates. S2: Pad the sparse data to the same size as the original radio map. S3: The sparsely sampled radio map is input into the designed lightweight Deep High-Resolution (DHR) module for image feature enhancement. S4: The enhanced image feature data is input into the improved generator module, and the Dual-Attention Multi-Scale Feature Fusion module (DAMF) and Transformer are combined to extract further feature information. S5: The additional conditional information and the real or generated radio map are input into the discriminator for true or false judgment. S6: Design a loss function to minimize the loss and update the generator and discriminator. S7: Repeat S1 to S6 until the loss value stabilizes.

2. The method for estimating a radio environment map based on a generative adversarial network according to claim 1, wherein: In step S1, the target area is divided into N×N grid units, and each K i Each small area contains a sensor. Each sensor receives the radio signal data of the area. These data are recorded in the form of signal power spectrum density (PSD). Specifically, the Kth i The sparse samples of the i-th region accepted by the sensor are (x i , p i ),in is the PSD of non-negative observations, is the corresponding two-dimensional coordinate. In addition, region i also has a corresponding city map The city map shows the building distribution information, where a value of 1 indicates that it is blocked by buildings. The sampling rate is set according to the actual scene, K i ≤N×N. Generally, a lower sampling rate can simulate a low density of measurement points, while a higher sampling rate is used for more accurate area coverage and signal capture.

3. The method for estimating a radio environment map based on a generative adversarial network according to claim 1, wherein: In step S2, a sparsely observed radio signal dataset is first obtained. These datasets are usually collected by sensors located in specific locations within a given area. The number of radio signal samples obtained is limited, and the area covered by these samples is much smaller than the entire target radio map area. In order to enable these sparse data to be input into the subsequent estimation module, they must be padded to the same size as the original radio map. Specifically, the sparse observations (x i , p i ) is zero-filled to fill the observation area to the same size as the original radio map, i.e. The interpolation result will not affect other features. Therefore, when the input feature has m distribution information, the feature is encoded into two channels, namely z i ={x′ i ,m i }, when there is no m distribution information, the feature is encoded as a channel, namely z i ={x′ i }, each channel size is N×N.

4. The method for estimating a radio environment map based on a generative adversarial network according to claim 1, wherein: In step S3, the zero-padded, sparsely sampled radio map data is first input into the designed lightweight DHR module. The DHR module is used to extract fine-grained features from the low-resolution radio map, particularly high-frequency details, thereby improving the quality of radio map reconstruction, especially signal recovery under sparse observation conditions. The DHR module utilizes a five-layer shallow convolutional architecture, significantly reducing computational complexity and memory usage compared to traditional 20-layer super-resolution networks. Specifically, except for the input and output layers, each layer of the DHR module, with the exception of the first and last layers, contains 64 convolutional kernels of size 3×3×64. Batch normalization and ReLU activation functions are used to improve training efficiency. These designs optimize the model training process, reducing network training time and increasing feature learning capabilities. In particular, the layer-by-layer residual connection mechanism preserves original feature information, avoids the vanishing gradient phenomenon, and promotes the efficient propagation of high-frequency information, ensuring that the network can extract richer details from low-resolution input data.

5. The method for estimating a radio environment map based on a generative adversarial network according to claim 1, wherein: In step S4, the processed data is input into an improved generator module, wherein the generator module adopts an encoder-decoder structure. The decoder uses skip connections and five layers of upsampling convolution blocks to gradually restore the resolution of the radio map and combine the features extracted by the encoder. The encoder of the generator maps the input data into a high-dimensional feature space through convolution blocks, and gradually reduces the spatial size through downsampling operations, while increasing the number of feature channels. To enhance the ability to extract signal features in complex environments, the encoder integrates a dual attention multi-scale feature fusion module (DAMF). The DAMF module combines void convolution and attention mechanism. DAMF extracts multi-scale features with different receptive fields, enhancing the ability to model local, regional, and global changes in the signal. The Position Attention Module (PAM) and Channel Attention Module (CAM) mechanisms are further introduced to dynamically adjust the weights of the spatial and channel dimensions of the features, thereby enhancing the response of key areas and important channels and effectively suppressing noise interference. The feature map processed by the DAMF block is input into the Transformer module. The Transformer module uses a self-attention mechanism to model long-range dependencies, addressing the limitations of traditional convolutional neural networks (CNNs) in modeling global features. The Transformer employs a multi-head self-attention mechanism to effectively integrate local information from sparse observation points with global information from areas obscured by buildings. Combined with position encoding, the Transformer maintains the spatial consistency of the radio map.

6. The method for estimating a radio environment map based on a generative adversarial network according to claim 1, wherein: In step S5, the generated radio map, along with additional conditional information, is input into the discriminator for authenticity verification. The discriminator utilizes a lightweight, stable, and efficient four-layer convolutional network architecture, effectively reducing computational complexity and optimizing the stability of generator training. The discriminator distinguishes between radio maps generated by the generator and those derived from real-world measurement data by performing a authenticity check on the input data. To enhance the discriminator's discriminative capabilities, a three-channel input mechanism was designed, consisting of a real or generated radio map and two channels of one-hot encoded prior information. The one-hot encoding provides regional geometric features and signal marker information, which helps the discriminator accurately determine the authenticity of input data in complex environments. Specifically, the input data is processed by a four-layer convolutional network, with each layer performing further feature extraction on the input. The discriminator finally outputs a one-dimensional probability value, which is used to judge whether the input radio map is real data or fake data generated by the generator. The output value is a value between 0 and 1. The closer the value is to 1, the better the radio map is. The greater the probability that the data is real data; the closer the value is to 0, the greater the probability that the data is generated data.

7. The method for estimating a radio environment map based on a generative adversarial network according to claim 1, wherein: In step S6, we used the following objective function to train the generator and discriminator: where p data (r, I) is the joint distribution together with the prior knowledge, Represents the unique hot encoding in the two channels. To optimize the training model, the maximum-minimum joint optimization problem is split into two sub-problems, and the generator G and the discriminator D are trained iteratively. Specifically, the loss function of the discriminator is: L D =-V(D,G) (2) The loss function used to optimize the generator is: To help the improved generator learn a feature distribution closer to the real radio map, we redesigned the generator's loss function. The generator loss combines global reconstruction loss, total variation loss (TV), gradient similarity loss (GS), and building-constrained loss (BC) to achieve efficient generator learning.

8. The method for estimating a radio environment map based on a generative adversarial network according to claim 1, wherein: In step S7, steps S1 to S6 are repeatedly performed until the loss values ​​of the generator and the discriminator tend to be stable and the training process is completed.

9. A radio environment map estimation method based on a generative adversarial network, characterized in that: include: The data acquisition module is used to collect the current radio map raw data according to different sampling rates. The data padding module is used to pad the sparse data to the same size as the original radio map. Feature enhancement module, which inputs the sparsely sampled radio map into the designed lightweight DHR for image feature enhancement. The generator module is used to input the processed data into the improved generator module, combining DAMF and Transformer. The discriminator module inputs the additional conditional information and the real or generated radio map into the discriminator for true or false judgment. The training optimization module is used to design a loss function that minimizes the loss and updates the generator and discriminator until the loss value stabilizes and the training process is completed.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the radio environment map estimation method based on the generative adversarial network according to any one of claims 1 to 8 is implemented.

Images