Electromagnetic spectrum map construction method for scene with undetectable area

Abstract

The invention discloses an electromagnetic spectrum map construction method for a scene with an undetectable area. The method comprises the following steps: acquiring spectrum data and geographical environment information of a target area sensing node with the scene with the undetectable area; generating an environment map of the target area; performing spatial grid discretization on the target area, creating a missing spectrum map image and a sampling position image of the target area, and creating a building map image; modeling a spectrum map missing problem as an image inpainting problem, and establishing a spectrum map completion model based on a partially convolutional UNet network; inputting the three-channel color spectrum map image, the sampling position image and the single-channel building map image into a spectrum map completion model, and training the spectrum map completion model; and processing an output result of the frequency spectrum map completion model, and outputting a complete frequency spectrum map. According to the invention, the reconstruction precision of the spectrum map can be effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of wireless communication technologies, and particularly relates to a method for constructing an electromagnetic spectrum map for a scenario with an undetectable area. Background Art

[0002] With the rapid development of wireless communication technologies, towards the era of the sixth-generation mobile communication network (6G), diverse wireless services have emerged continuously. Against this background, the accurate construction of a spectrum map has become particularly crucial. It reflects the accurate distribution of spectrum resources in the radio environment, can effectively support the dynamic sharing and intelligent management and control of spectrum resources, and is of vital importance for improving the utilization efficiency of spectrum resources.

[0003] Existing spectrum map construction methods can be divided into model-based methods and model-free (data-driven) methods. The model-based spectrum map construction method is based on a priori models, mainly using radio propagation models. However, it is sensitive to the selection of the propagation model and its related parameters, and may be difficult to consider specific geographical environment factors, such as shadows or obstacles. The model-free spectrum map construction method can reconstruct a radio map only based on measurement data without relying on a specific radio propagation model.

[0004] Publication No. CN116578660A discloses a method for constructing an electromagnetic target situation map based on sparse data. This method improves the model-based method and considers the influence of unknown environments on spectrum signal strength, the base mismatch problem of target positions, and the mutual interference problem between adjacent targets. Publication No. CN115205481A discloses a method for constructing a spectrum map based on a graph neural network, which uses a graph neural network to combine geographical location information and signal strength data to construct a high-precision spectrum map. However, this method assumes that the measurement data is uniformly distributed and can perform spectrum monitoring for all spatial positions. Most of the existing spectrum map construction technologies consider scenarios where all regions are detectable and the sensing data is uniform. However, in actual scenarios, due to security considerations or physical limitations, unauthorized spectrum devices are usually prohibited from entering certain inaccessible areas, resulting in blank areas on the radio map. Therefore, traditional spectrum map reconstruction methods are difficult to be applicable to the construction of spectrum maps in scenarios with non-uniformly distributed measurements and inaccessible areas, and there is an urgent need to develop new spectrum map construction methods.

[0005] In complex scenarios, such as in the case of inaccessible areas, the model adaptability is limited. Therefore, how to use deep learning methods to achieve high-precision spectrum map construction under the coexistence of sparse data sampling and inaccessible areas is an urgent problem to be solved. Summary of the Invention

[0006] The object of the present invention is to provide a method for constructing an electromagnetic spectrum map for a scenario with an undetectable area, which is particularly suitable for a complex environment with an undetectable area. By using partial convolution and combining with the dynamic update of the sampling position image, the limited spectrum data in the accessible area is fully utilized to gradually update the missing part of the accessible area and the spectrum map of the inaccessible area. At the same time, by fusing geographical environment information, the reconstruction accuracy of the spectrum map is further improved.

[0007] To achieve the above technical object, the technical solution adopted by the present invention is as follows:

[0008] The present invention discloses a method for constructing an electromagnetic spectrum map for a scenario with an undetectable area, and the method includes the following steps:

[0009] S1, collecting spectrum data and geographical environment information of the target area perception nodes in the scenario with an undetectable area;

[0010] S2, preprocessing the collected spectrum data, and generating an environmental map of the target area based on the collected geographical environment information;

[0011] S3, discretizing the target area into spatial grids, creating a missing spectrum map image and a sampling position image of the target area according to the preprocessed spectrum data, and creating a building map image according to the generated environmental map;

[0012] S4, modeling the spectrum map missing problem as an image inpainting problem, based on the UNet structure of partial convolution, and combining with the dynamic update of the sampling position image, establishing a spectrum map completion model; the input of the spectrum map completion model is a three-channel color sampled spectrum map image and a single-channel building map image, and the output is the completed spectrum map;

[0013] S5, collecting complete spectrum map images, building map images and randomly generated sampling position images at different base station positions to form an original data set, dividing the original data set according to a ratio to obtain a training set, a validation set and a test set; using the training set data and the validation set data to train the spectrum map completion model;

[0014] S6, inputting the test set data into the best-trained spectrum map completion model, and outputting a complete spectrum map.

[0015] In step S2, the collected spectrum data is preprocessed by normalization, and the spectrum map measurement value is standardized between 0 and 1.

[0016] Step S3 further includes:

[0017] The target area is equally divided into a discrete grid of size S×N at fixed intervals along the latitude and longitude directions, and the target area is equally divided into S×N square blocks of the same size;

[0018] Create the missing spectral map image:

[0019]

[0020] where are the latitude and longitude coordinates, and Ω is the corresponding received signal strength; if the grid I=(i,j) contains one sampling value, then assign [Ω k i,j to the received signal strength collected at the corresponding grid point; if the grid I=(i,j) contains M sampling values, then assign [Ω k i,j to the average value of the M received signal strengths; if the grid I=(i,j) has no sampling values, then assign [Ω k i,j to 0; the missing spectral map at frequency f k is expressed as: as:

[0021]

[0022] where Ψ(f k ,c i,j ) represents the received signal strength at the grid I=(i,j), the subscript m represents the m-th sampling value, 1<m≤M, represents the missing spectral map with the transmission frequency of f k , represents the spectral data at the grid I=(i,j) when the transmission frequency is f k .

[0023] Furthermore, in step S3, the sampling position image is:

[0024]

[0025] The building map image is:

[0026]

[0027] where B i,j represents the value of the building map at the grid I=(i,j).

[0028] ​​​Further, in step S4, the spectrum map completion model uses partial convolution combined with sampling position update operation to gradually complete the missing spectrum map, and finally complete the construction of the electromagnetic spectrum map; the expression of the spectrum map completion model is:

[0029]

[0030] where g(·) is the spectrum map completion model, is the complete spectrum map output by the completion model.

[0031] Further, in step S4, the partial convolution operation is as follows:

[0032]

[0033] where W is the convolution kernel, (·) T represents transpose, X is the feature matrix of the current convolution (sliding) window, A is the corresponding binary mask matrix, and ⊙ represents element-wise multiplication; X in the first layer is a part of the real spectrum map Ψ, and A is a part of the sampling position image M; b is the bias term of the convolution operation, 1 is a matrix with the same size as the convolution kernel and all its elements are 1, and the output x' is a feature value for the next layer of convolution operation.

[0034] Further, in step S4, the dynamic update expression of the sampling position image is as follows:

[0035]

[0036] where when the feature matrix of the mask matrix A contains a value of 1, it means that the output of the corresponding partial convolution operation is valid, and this position is marked as valid as the input for the next layer; after each convolution operation, the sampling position image is dynamically updated according to the output features, so that the network gradually expands the valid area in subsequent convolutions to complete the missing area.

[0037] Further, in step S5, the training process of the spectrum map completion model includes:

[0038] S51, multiply the complete spectrum map and the matrix corresponding to the randomly generated sampling position image element by element to obtain a three-channel sampled spectrum map image, and form the training set data together with the single-channel building map image;

[0039] S52, initialize the network training iteration number epoch to 1, and initialize the maximum iteration number, learning rate, and training optimizer;

[0040] S53. Input the training set data in batches into the UNet network with partial convolution for training. Backpropagate the training error of each batch to optimize the network parameters. When all batches of data in the training data have been backpropagated, one iteration is completed.

[0041] S54. Determine whether the current training error is smaller than the previous one. If so, update the network model and save the parameters of this network model. If not, do not update the network model.

[0042] S55. Determine whether the network training is completed: Check whether the current epoch has reached the set maximum epoch. If so, output the trained spectral map completion model. If not, increment the epoch by 1 and return to step S52 to continue network training.

[0043] Furthermore, in step S5, the error loss function l of the spectral map completion model total is:

[0044]

[0045] where λ1, λ2, λ3, λ4, and λ5 are the corresponding loss weights; l s and l u-s are the sampling value loss functions:

[0046] l u-s = ||(1 - M) ⊙ (Ψ o - Ψ)||1

[0047] l s = ||M ⊙ (Ψ o - Ψ)‖1

[0048] In the formula, ⊙ is element-wise multiplication, Ψ o is the spectral map generated by the spectral map completion model, Ψ is the real spectral map, and ||·||1 represents the L1 distance;

[0049] l pe is the perceptual loss function:

[0050]

[0051] In the formula, Q is the number of activation layers, Y q represents the activation map of the q-th selected layer extracted from the pre-trained VGG-16 model on ImageNet. The activation maps extracted from the pre-trained VGG-16 model on ImageNet are used to calculate the differences between the input image and the target image on these feature layers;

[0052] and are the style loss functions:

[0053]

[0054] In the formula, E q is a factor for normalizing the activation map of the q-th selected layer;

[0055] l tv is the total variation loss function:

[0056]

[0057] In the formula, (x, y) are the position coordinates corresponding to the pixels of the spectral ground image, and P represents the area after expanding the missing area by 1 pixel.

[0058] Furthermore, the following formula is used to process the output result of the spectral map completion model to output the complete spectral map:

[0059]

[0060] In the formula, is the sampling position image, and Ψ o is the spectral map generated by the spectral map completion model.

[0061] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0062] The electromagnetic spectral map construction method for scenarios with non-detectable regions of the present invention, aiming at the spectral missing problem, proposes to model it as an image inpainting problem. To better distinguish the sampled area and the non-sampled area, partial convolution is used to replace ordinary convolution. The present invention is based on the neural network method of partial convolution and comprehensively considers the geographical environment information to assist in the construction of the spectral map of the target area, can quickly realize the construction of the spectral map, and significantly improves the construction accuracy of the spectral map. Description of the Drawings

[0063] Figure 1 is the flow chart of the electromagnetic spectral map construction method for scenarios with non-detectable regions of the present invention;

[0064] Figure 2 is the work flow chart of the framework of the present invention and the schematic diagram of the corresponding network structure G-PcNet and related parameters;

[0065] Figure 3 is the comparison diagram of the reconstruction error of the increase in the number of large-area missing region blocks in the 10% sampling rate scenario between the present invention and other methods;

[0066] Figure 4 is the comparison diagram of the reconstruction error of different sampling rates between the method of the present invention and other methods in the case of the same large-area missing region;

[0067] Figure 5 This is a comparison chart of the visualization effects of the reconstructed spectral maps of the method of the present invention and other methods at a 10% sampling rate, with a large missing area of 50×50m 2 ;

[0068] Figure 6 is Figure 5 the corresponding normalized MSE result schematic diagram. Detailed implementation manners

[0069] The embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0070] Although the steps in the present invention are arranged with reference numerals, they are not used to limit the order of the steps. Unless the order of the steps is clearly stated or the execution of a certain step requires other steps as a basis, the relative order of the steps can be adjusted.

[0071] Combined with Figure 1 the specific steps of the method for constructing an environmental spectral map considering environmental information and facing an environment with undetectable areas of the present invention are described as follows:

[0072] Step 1, collect the spectral data of the sensing nodes in the target area with undetectable areas, and collect the geographical environment information of the target area;

[0073] Step 1.1, collect the spectral data of the target area with undetectable areas;

[0074] The first step is to clarify and set the target area where spectral data needs to be collected;

[0075] The second step is to collect spectral data in the target area using high-precision sensing devices according to the predetermined sampling ratio requirements;

[0076] The third step is to upload the collected spectral data to the central database for storage.

[0077] In the embodiments of the present invention, due to cost constraints in deployment, the number of sensing nodes is often limited and unevenly distributed. In addition, due to security considerations and some physical limitations, unauthorized spectral devices are usually prohibited from entering certain inaccessible areas, resulting in the inability to collect data in these areas. Eventually, only a discrete, unevenly distributed spectral map with large blank areas can be obtained.

[0078] Step 1.2, collect the geographical environment information of the target area;

[0079] Collect detailed geographical environment information of the area using Geographic Information System (GIS). Based on the collected data, generate an environmental map of the target area and upload the environmental map of the target area to the central database.

[0080] Step 2, data preprocessing: Discretize the target area into spatial grids to obtain a missing spectrum map image; process the environmental map of the target area to create a building map image.

[0081] Step 2.1: Discretize the target area into spatial grids to obtain a missing spectrum map image.

[0082] First step: Discretize the target area into grids.

[0083] Divide the target area into square grids with equal length and width at a fixed interval. The target area is thus divided into S×N discrete grids.

[0084] After discretizing the target area into grids, a grid may contain one or more sensing nodes. However, due to cost constraints in deployment, most grids do not contain sensing nodes. Additionally, due to security considerations and some physical limitations, unauthorized spectrum devices are usually prohibited from entering certain inaccessible areas, resulting in the inability to deploy sensing nodes in these areas. Therefore, most grids do not contain sensing nodes, and only some grids contain multiple sensing nodes. Based on this, create a missing spectrum map image:

[0085] If grid I=(i,j) contains one sampling value, then assign [Ω k i,j the received signal strength collected at the corresponding grid point; if grid I=(i,j) contains M sampling values, then assign [Ω k i,j the average of the M received signal strengths; if grid I=(i,j) does not contain a sampling value, then assign [Ω k i,j 0.

[0086] Correspondingly, the missing spectrum map k at frequency f is expressed as:

[0087]

[0088] where Ψ(f k ,c i,j ) represents the received signal strength at grid I=(i,j), the subscript m represents the m-th sampling value, 1 < m ≤ M, represents the missing spectrum map with transmission frequency f k , ​​​Indicates the transmission frequency is f k The spectral data at grid I=(i,j) when the frequency is f.

[0089] Step 2.2: Extract the collected geographical environment information, create a building map image and input it into the network to assist in the construction of the spectral map;

[0090] The first step: Create a binary building map

[0091]

[0092] Among them, B i,j Represents the value of the building map at grid I=(i,j).

[0093] In the embodiment of the present invention, considering that due to security considerations and some physical limitations, unauthorized spectrum devices are usually prohibited from entering certain inaccessible areas, which will cause large blank areas in the missing spectral map. To enable the neural network to perform better inference, the present invention considers binary building map information to assist in the construction of the spectral map.

[0094] Step 3: Establish a spectral map completion model;

[0095] Model the spectral map missing problem as an image inpainting problem:

[0096]

[0097] Among them, g(·) is the spectral map completion model, Is the complete spectral map output by the completion model.

[0098] Step 4: Input the three-channel sampled spectral map image and the single-channel building map image into the network model, and use the training set data and the validation set data to train the network model; during the network training process, the real spectral map is obtained from the complete spectral map, which is only used to calculate the loss and guide the network learning, but in the inference stage, only the sampled area has real values, and for the non-sampled area, the network must complete these areas through the context relationship and the inference ability of the model, so as to output the complete spectral map. The training process includes the following steps:

[0099] Step (1), as Figure 2 Shown, input the three-channel sampled spectral map image (the sampled spectral map image in the training stage is obtained by element-wise multiplication of the complete spectral map and the randomly generated sampling position image), and the single-channel building map image into the UNet network G-PcNet using partial convolution;

[0100] Step (2): Initialize the network training parameters. Initialize the network training iteration number epoch to 1, the maximum iteration number to 50, the learning rate to 0.0001, and use the Adam optimization algorithm as the network training optimizer.

[0101] Step (3): Input the training data into the network in batches for training, and backpropagate the training error of each batch to optimize the network parameters. The error loss function l total is calculated as follows:

[0102]

[0103] where l s and l u-s are the sampling value loss functions:

[0104] l u-s = ||(1 - M) ⊙ (Ψ o - Ψ)||1 (5)

[0105] l s = ||M ⊙ (Ψ o - Ψ)||1 (6)

[0106] where ⊙ is the element-wise multiplication, Ψ o is the spectral map generated by the network, and ||·||1 represents the L1 (absolute value) distance; l pe is the perceptual loss function:

[0107]

[0108] Y q represents the activation map of the q-th selected layer extracted from the pre-trained VGG-16 model on ImageNet. and are the style loss functions:

[0109]

[0110] E q is the factor used to normalize the activation map of the q-th selected layer. l tv is the total variation loss function:

[0111]

[0112] P represents the region after expanding the missing region by 1 pixel. λ1, λ2, λ3, λ4, and λ5 are the corresponding loss weights, which are finally 1, 6, 0.1, 0.05, and 120 respectively.

[0113] When all batches of data in the training data have completed backpropagation, it is 1 epoch;

[0114] Step (4), save the optimal network model: Each time the backpropagation in step (3) is completed, save the network model once. Determine whether the current training error is smaller than the previous one. If so, update the network model and save the parameters of this network model. If not, do not update the network model.

[0115] Step (5), determine whether the network training is completed: Determine whether the current epoch has reached the set maximum epoch. If so, proceed to step 6. If not, increment the epoch by 1 and return to step (2) to continue the network training.

[0116] Step 6, input the test set data into the best-trained network model, process the output results, and output the complete spectrum map.

[0117] Process the output results and output the complete spectrum map

[0118] The effects of the present invention will be further described below in conjunction with simulation experiments.

[0119] 1. Simulation conditions and parameter settings:

[0120] The simulation experiment of the present invention is carried out on a simulation platform of Python 3.11.3 and Pytorch 2.0.1. The server CPU model is an Intel Core i9-13900K processor, equipped with an independent graphics card of model NVIDIA GeForce RTX4090.

[0121] The maximum number of iterations for network training is 50, the learning rate is 0.001, the Adam optimization algorithm is selected as the network training optimizer, and the data volume of each batch is 12.

[0122] 2. Simulation content:

[0123] For convenience, the spectrum map measurement values are further normalized between 0 and 1. Since the original spectrum map and the normalized spectrum map have exactly the same pattern, the normalized spectrum map can be complemented and transformed. Regarding the complementation performance of the spectrum map, the mean square error MSE is used for measurement. The smaller the MSE, the better the complementation effect.

[0124]

[0125] Among them, M represents the total number of samples, m represents the sample serial number, 1 ≤ m ≤ M; ||·|| F represents the tensor Frobenius norm, Ψ m represents the missing spectrum map data, represents the true spectrum map data, and g(·) represents the spectrum map complementation model.

[0126] The present invention comprehensively compares five methods, namely the neural network method PcNet based on partial convolution, the classical algorithms NW and FDGM in the field of image inpainting, and the representative methods Kriging and IDW in the field of spectral map completion. To achieve a fair and intuitive comparison, the six MSE results in a single application scenario are normalized for quantitative analysis of the advantages and disadvantages of different methods. Specifically, the minimum value among the six MSE results is regarded as the benchmark, normalized to 0, the maximum value is normalized to 1, and the remaining values are calculated according to the following formula:

[0127]

[0128] where MSE norm,i is the normalized MSE of the i-th method in the current scenario, MSE i is the original MSE of the i-th method in the current scenario, MSE min and MSE max are the minimum and maximum values of the MSE of all methods in this scenario, respectively. Through this normalization operation, the MSE results of each scenario are mapped to the interval [0,1], thus eliminating the interference of different dimensions and ensuring the fairness and consistency of the comparison.

[0129] Figure 3 is the comparison chart of the reconstruction errors with and without considering environmental information under different large-area missing region sizes in the present invention. The considered scenarios are: the target region size is 512×512m 2 , the sampling rate is 10%, and the size of the inaccessible region is 50×50m 2 , and the completion performance when the number ranges from 1 to 4. Figure 4 In, the abscissa represents the number of inaccessible regions, and the ordinate is the normalized MSE. It can be seen from the Figure 4 comparison that the proposed completion method considering environmental information is superior to other methods.

[0130] Figure 4 is the performance comparison chart with and without considering environmental information under different sampling rate scenarios in the present invention. The considered scenarios are: the target region size is 512×512m 2 , where the size of the inaccessible region is 50×50m 2 , and the completion performance when the sampling rates are 1%, 5%, 10%, and 100% respectively. Figure 3 In, the abscissa represents the sampling rate, and the ordinate is the normalized MSE. It can be seen from the Figure 3 comparison that the proposed completion method considering environmental information is superior to other methods.

[0131] See Figure 5 andFigure 6 As can be seen from the visualization effect diagram, the method using deep learning is significantly better than the ordinary method. Regarding the completion performance of inaccessible areas, considering environmental information is better than not considering environmental information.

[0132] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of the present application can be implemented in various computer languages. For example, object-oriented programming languages such as Java and interpreted scripting languages such as JavaScript.

[0133] The present application is described with reference to the flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions run by the processor of the computer or other programmable data processing devices generate a device for implementing the specified functions in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0134] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory generate a manufactured article including an instruction device, and the instruction device implements the specified functions in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0135] These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are run on the computer or other programmable device to generate a computer-implemented process. Thus, the instructions running on the computer or other programmable device provide steps for implementing the specified functions in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0136] Although the preferred embodiments of the present application have been described, those skilled in the art can make additional changes and modifications once they learn the basic creative concept. Therefore, the appended claims are intended to be construed as including the preferred embodiments as well as all changes and modifications falling within the scope of the present application.

[0137] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalent technologies, the present application is also intended to include these modifications and variations.

Claims

1. A method for constructing an electromagnetic spectrum map for a scene with an undetectable area, characterized in that: The method comprises the following steps: S1, collects spectrum data and geographic environment information of the sensing nodes in the target area with undetectable area scenes; S2, preprocessing the collected spectrum data, and generating an environmental map of the target area based on the collected geographical environment information; S3, discretizing the target area into a spatial grid, creating a missing spectrum map image and a sampling position image of the target area according to the preprocessed spectrum data, and creating a building map image according to the generated environment map; S4, the spectrum map missing problem is modeled as an image restoration problem. Based on the partially convolutional UNet structure and combined with the dynamic update of the sampling position image, a spectrum map completion model is established; the input of the spectrum map completion model is a three-channel color sampling spectrum map image and a single-channel building map image, and the output is the completed spectrum map; S5, collecting complete spectrum map images, building map images and randomly generated sampling location images at different base station locations to form an original data set, dividing the original data set in proportion to obtain a training set, a validation set and a test set; using the training set data and the validation set data to train a spectrum map completion model; S6, input the test set data into the best trained spectrum map completion model and output the complete spectrum map.

2. The method for constructing an electromagnetic spectrum map for a scene with an undetectable area according to claim 1, characterized in that: In step S2, the collected spectrum data is normalized and preprocessed to standardize the spectrum map measurement value between 0 and 1.

3. The method for constructing an electromagnetic spectrum map for a scene with an undetectable area according to claim 1, characterized in that: Step S3 further comprises: The target area is equally divided into discrete grids of size S×N at fixed intervals along the latitude and longitude directions, and the target area is equally divided into S×N square blocks of the same size; Create the missing spectrum map image: In the formula, is the longitude and latitude coordinates, Ω is the corresponding received signal strength; if the grid I=(i,j) contains a sample value, then [Ω k ] i,j Assign the received signal strength collected at the corresponding grid point; if the grid I = (i, j) contains M sampling values, then [Ω k ] i,j Assign the value to the average value of M received signal strengths; if there is no sample value in the grid I = (i, j), then [Ω k ] i,j Assign value 0; frequency f k Missing spectrum map under It is expressed as: In the formula, Ψ(f k ,c i,j ) represents the received signal strength at the grid I = (i, j), the subscript m represents the mth sampling value, 1<m≤M, Indicates that the transmission frequency is f k The missing spectrum map, Indicates that the transmission frequency is f k The spectrum data at the time grid I = (i, j).

4. The method for constructing an electromagnetic spectrum map for a scene with an undetectable area according to claim 1, characterized in that: In step S3, the sampling position image for: Building map image for: In the formula, B i,j Represents the value of the building map at grid I = (i, j).

5. The method for constructing an electromagnetic spectrum map for a scene with an undetectable area according to claim 1, characterized in that: In step S4, the spectrum map completion model uses partial convolution combined with sampling position update operation to gradually complete the missing spectrum map and finally complete the electromagnetic spectrum map construction; the expression of the spectrum map completion model is: Among them, g(·) is the spectrum map completion model, Complete spectrum map output by the completion model.

6. The method for constructing an electromagnetic spectrum map for a scene with an undetectable area according to claim 1, characterized in that: In step S4, the partial convolution operation is as follows: Where W is the convolution kernel, (·) T represents transpose, X is the feature matrix of the current convolution (sliding) window, A is the corresponding binary mask matrix, ⊙ represents element-by-element multiplication; X of the first layer is part of the real spectrum map Ψ, A is part of the sampling position image M; b is the bias term of the convolution operation, 1 is a matrix of the same size as the convolution kernel, all elements of which are 1, and the output x' is an eigenvalue of the next layer of convolution operation.

7. The method for constructing an electromagnetic spectrum map for a scene with an undetectable area according to claim 1, characterized in that: In step S4, the dynamic update expression of the sampling position image is as follows: Among them, when the feature matrix of the mask matrix A contains a value of 1, it means that the output of the corresponding partial convolution operation is a valid value, and the position is marked as valid as the input of the next layer; after each convolution operation, the sampling position image is dynamically updated according to the output features, so that the network can gradually expand the valid area in subsequent convolutions and complete the missing area.

8. The method for constructing an electromagnetic spectrum map for a scene with an undetectable area according to claim 1, characterized in that: In step S5, the training process of the spectrum map completion model includes: S51, performing element-by-element multiplication of the complete spectrum map and the matrix corresponding to the randomly generated sampling position image to obtain a three-channel sampling spectrum map image, which is combined with the single-channel building map image to form training set data; S52, initialize the network training iteration number epoch to 1, initialize the maximum number of iterations, learning rate and training optimizer; S53, input the training set data in batches into the UNet network using partial convolution for training, back-propagate the training error of each batch to optimize the network parameters, and complete one iteration when all batches of data in the training data are back-propagated; S54, determining whether the current training error is smaller than the previous one, if so, updating the network model and saving the network model parameters, if not, not updating the network model; S55, judging whether the network training is completed: judging whether the current epoch reaches the set maximum epoch, if so, outputting the trained spectrum map completion model, if not, adding 1 to the epoch, returning to step S52, and continuing network training.

9. The method for constructing an electromagnetic spectrum map for a scene with an undetectable area according to claim 1, characterized in that: In step S5, the error loss function l of the spectrum map completion model total for: Among them, λ1, λ2, λ3, λ4 and λ5 are the corresponding loss weights; l s and l u-s is the sampling value loss function: l u-s =||(1-M)⊙(Ψ o -Ψ)||1 l s =||M⊙(Ψ o -Ψ)||1 Where ⊙ is the element-wise multiplication, Ψ o is the spectrum map generated by the spectrum map completion model, Ψ is the real spectrum map, and ||·‖1 represents the L1 distance; l pe is the perceptual loss function: Where Q is the number of activated layers, Y q represents the activation map of the qth selected layer extracted from the VGG-16 model pre-trained on ImageNet, using the activation map extracted from the VGG-16 model pre-trained on ImageNet to calculate the difference between the input image and the target image on these feature layers; and is the style loss function: In the formula, E q is the factor used to normalize the activation map of the qth selected layer; l tv is the total variation loss function: Where (x, y) is the position coordinate corresponding to the pixel of the spectrum map, and P represents the area after the missing area is expanded by 1 pixel.

10. The method for constructing an electromagnetic spectrum map for a scene with an undetectable area according to claim 1, characterized in that: The following formula is used to process the output result of the spectrum map completion model and output a complete spectrum map: In the formula, is the sampling position image, Ψ o is the spectral map generated by the spectral map completion model.

Images