Method for predicting parameters of a wireless channel based on geometric geographic information

Abstract

The application provides a wireless channel characteristic parameter prediction method based on geometric geographic information, which converts a spatial channel parameter prediction problem into an image pixel value completion problem, constructs a scattering environment feature map by using known environment features, and realizes prediction of channel characteristic parameters of a three-dimensional scene by combining a deep learning method. The application comprehensively considers the internal connection between channel characteristics and scattering environment characteristics, integrates scene information into a network training process, and realizes accurate and effective prediction of channel characteristics in a wireless communication scene by using a scene and data double-driving method. Meanwhile, the application proposes an improved GAIN prediction network framework, the network uses multi-channel feature data composed of a scattering environment feature map and partially sampled channel characteristics as input, and can convert channel characteristic parameter prediction at different positions into image pixel value prediction.

Description

Technical Field

[0001] This invention belongs to the field of wireless information transmission, specifically relating to a method for predicting wireless channel characteristic parameters based on geometric geographic information, and is aimed at applications of wireless channel characteristic parameter prediction using scene information and channel characteristic parameters. Background Technology

[0002] The rapid development of 5G and 5G+ mobile communication technologies has placed new demands on communication reliability and performance. Channel modeling, as a crucial component of wireless communication systems, is of great significance for system design, improving communication quality, and evaluating communication performance. Wireless channel characteristics are determined by the physical propagation environment of wireless communication and the interacting scatterers within it; different scattering scenarios produce different channel characteristics. Using scattering environment characteristics as prior information to predict channel states allows for the development of environment-based adaptive communication technologies, such as transmission power loss mitigation, intelligent decision-making, intelligent resource allocation and scheduling, and motion trajectory planning.

[0003] In recent years, machine learning has developed rapidly and has been widely applied in wireless communication and channel modeling. Machine learning-based methods can autonomously analyze data features and learn the complex relationships between potential inputs and outputs, thereby predicting data trends. Existing machine learning-based parameter prediction methods typically only obtain the nonlinear relationship between coordinates and one or a few parameters, without utilizing prior information about the propagation environment to assist in prediction. Therefore, how to effectively utilize known scattering environment features and a small number of channel feature parameters sampled from known locations, combined with deep learning methods, to achieve fast and accurate prediction of channel feature parameters across all scenarios is a key research problem. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method for predicting wireless channel feature parameters based on geometric geographic information. This method transforms the spatial channel parameter prediction problem into an image pixel value completion problem, constructs a scattering environment feature map using known environmental features, and combines deep learning methods to predict the channel feature parameters of a three-dimensional scene.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A method for predicting wireless channel characteristic parameters based on geometric geographic information, characterized by comprising:

[0007] Step 1: Input initial parameters, including scene parameters and network parameters;

[0008] Step 2: Reconstruct the scene of the area to be tested based on the input initial parameters;

[0009] Step 3: In the reconstructed test area scene, based on the input sampling point position, the channel impulse response at the corresponding position is obtained by measurement at different height layers, and the channel feature parameters contained in the channel impulse response are extracted.

[0010] Step 4: Based on the channel feature parameters sampled in Step 3, construct a generative adversarial completion network for predicting channel feature parameters at different height levels;

[0011] Step 5: Vertically divide the channel feature parameters sampled in Step 3 into layers along the direction perpendicular to the ground to obtain channel feature parameters of different vertical layers. Based on the channel feature parameters of different vertical layers, construct a generative adversarial completion network for predicting the channel feature parameters of different vertical layers.

[0012] Step 6: Based on the generative adversarial completion networks with different height layers and different vertical layers constructed in Steps 4 and 5, predict the channel feature parameters of the location coordinates to be predicted in the scene.

[0013] To optimize the above technical solution, the specific measures also include:

[0014] Furthermore, in the first step, the scene parameters include the length, width, height, height layer grid resolution, vertical layer grid resolution, and sampling point position of the region to be tested; the network parameters include the number of network training iterations for the height layer and vertical layer, the error threshold, the maximum number of iterations not exceeding the error threshold, the generator learning rate, and the discriminator learning rate.

[0015] Furthermore, in the second step, the scene reconstruction of the area to be tested is performed as follows:

[0016] The channel impulse response measured at different heights is horizontally layered. Each height layer is then meshed according to a height layer grid resolution r1, resulting in an M×N resolution scatterer occlusion mask matrix M and a scattering environment feature map E corresponding to each height layer h. The scatterer occlusion mask matrix M uses 1 and 0 to represent areas covered and uncovered by buildings, respectively, and the scattering environment feature map E includes a scatterer height feature map E' describing the scatterer height. h and scatterer type feature map E used to describe the location of scatterer types c .

[0017] Furthermore, the scatterer height feature map E h The height values ​​in the graph are restricted to the range [0, 1] by normalization, and the calculation method is as follows:

[0018]

[0019] Where (i, j) represents the coordinates of the pixel. These are the original height and normalized height at point (i, j), respectively;

[0020] The scatterer type characteristic diagram E c The three types of scatterers—ground, buildings, and vegetation—are represented by a classification label of 0-2. An image channel containing only that type is created for each type. The image in each channel consists of 0s and 1s, where 1s represent the location distribution of the type in the image.

[0021] Furthermore, the third step is as follows:

[0022] Based on the sampling point location P, the corresponding sampling matrix S is obtained at each height layer. The channel impulse response of the receiver at the corresponding location is obtained by measurement. By extracting the channel feature parameters contained in the channel impulse response, a channel parameter feature map is constructed. Where K = 1, 2, and 3 correspond to path loss, root mean square delay spread, and root mean square angle spread, respectively.

[0023] Furthermore, the fourth step specifically includes the following steps:

[0024] S4.1: Randomly initialize the generator parameters θ of the network G and discriminator parameters θ D ;

[0025] S4.2: Optimize the training of the generator G network, as follows:

[0026] The generator G is based on the scattering environment feature map E and the channel parameter feature map including partial sampling. As input, a channel feature image with all pixels padded is generated through training, and the calculation method is as follows:

[0027]

[0028] Where ⊙ represents the Hadamard product, the generator's output image. Multiply by M to obtain the channel feature prediction map X at the effective location outside the building. g ;

[0029] Loss function of generator G as follows:

[0030]

[0031] Where, m ij Let D(·) represent the (i, j)th element of M. ij Let λ represent the (i, j)th element of the decision matrix P output by the discriminator D, and let λ represent the weight value used to balance the two parts of the loss.

[0032] The generator G uses a stochastic gradient optimizer to update its parameters in order to minimize the negative value of the objective function. The calculation method is as follows:

[0033]

[0034] in, It is the updated generator parameter, α G It is the generator learning rate. This represents the negative gradient value of the generator G loss function;

[0035] S4.3: Optimize the training of the discriminator D network, as follows:

[0036] Input of discriminator D It is a channel parameter feature map including partial sampling. The calculation method is as follows, combining the unsampled positions completed by generator G:

[0037]

[0038] The discriminator D outputs a discrimination matrix P, which is the result of calculating the binary cross-entropy loss for the matrix elements corresponding to the valid locations outside the building.

[0039] Loss function of discriminator D as follows:

[0040]

[0041] Among them, s ij This represents the (i, j)th element of S;

[0042] Discriminator D uses a stochastic gradient optimizer to update its parameters in order to minimize the negative value of the objective function. The calculation method is as follows:

[0043]

[0044] in, These are the updated discriminator parameters, α D It is the discriminator learning rate. This represents the negative gradient value of the loss function of the discriminator D;

[0045] S4.4: Repeat the training process of S4.2 and S4.3 until the iteration condition is met, then the network training ends and a generative adversarial completion network is obtained for predicting channel feature parameters at different height levels.

[0046] Furthermore, the fifth step specifically includes the following steps:

[0047] S5.1: Channel parameter feature map of the third sampling step The data is vertically layered according to a vertical grid resolution of r² along a direction perpendicular to the ground, and channel feature maps of different vertical layers are obtained. And the corresponding sampling matrix S′, where K=1,2,3 correspond to path loss, root mean square delay spread and root mean square angle spread respectively. At the same time, combined with the scene of the test area reconstructed in the second step, the vertical P×M resolution scatterer occlusion mask matrix M′ and scattering environment feature map E′ are obtained.

[0048] S5.2: Randomly initialize the generator parameters θ of the network G′ and discriminator parameters θ D′ ;

[0049] S5.3: Optimization of generator G′ and discriminator D′ network training:

[0050] 1) The process of training and optimizing the generator G′ network is as follows:

[0051] The generator G′ is based on the scattering environment feature map E′ and the channel parameter feature map including partial sampling. As input to the network, a channel feature image with all pixels padded is generated during training. The calculation method is as follows:

[0052]

[0053] Where ⊙ represents the Hadamard product, the generator's output image. Multiplying by M′ yields the channel feature prediction map X at the effective location outside the building. g ′;

[0054] Loss function of generator G′ as follows:

[0055]

[0056] Where, m ij Let ' represent the (i, j)th element of M', and D'(·) ij λ represents the (i, j)th element of the decision matrix P′ output by the discriminator D′, and λ′ represents the weight value used to balance the two parts of the loss;

[0057] The generator G′ uses a stochastic gradient optimizer to update its parameters in order to minimize the negative value of the objective function. The calculation method is as follows:

[0058]

[0059] in, It is the updated generator parameter, α G′ It is the generator learning rate. This represents the negative gradient value of the generator G′ loss function;

[0060] 2) The process of training and optimizing the discriminator D′ network is as follows:

[0061] Input of discriminator D′ It is a channel parameter feature map including partial sampling. The calculation method, combining the unsampled positions filled by generator G′, is as follows:

[0062]

[0063] The discriminator D′ outputs a discrimination matrix P′, which is a binary cross-entropy loss calculated from the matrix elements corresponding to the valid locations outside the building.

[0064] Loss function of discriminator D′ as follows:

[0065]

[0066] Among them, s ij ' represents the (i, j)th element of S';

[0067] The discriminator D′ uses a stochastic gradient optimizer to update its parameters in order to minimize the negative value of the objective function. The calculation method is as follows:

[0068]

[0069] in, These are the updated discriminator parameters, α D′ It is the discriminator learning rate. This represents the negative gradient value of the loss function of the discriminator D′;

[0070] S5.4: Repeat the training process of S5.3 until the iteration condition is met to obtain a generative adversarial completion network for predicting channel feature parameters of different vertical layers.

[0071] Furthermore, the iteration condition is that the number of training iterations i of the network reaches a set number, or the network loss function value does not exceed the error threshold ε for l consecutive iterations, where l is the maximum number of iterations that does not exceed the error threshold.

[0072] Furthermore, the present invention also proposes a computer-readable storage medium storing a computer program, characterized in that the computer program enables a computer to execute the wireless channel characteristic parameter prediction method based on geometric geographic information as described above.

[0073] Furthermore, the present invention also proposes an electronic device, characterized in that it includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the wireless channel feature parameter prediction method based on geometric geographic information as described above.

[0074] The beneficial effects of this invention are:

[0075] 1. This invention proposes a wireless channel feature parameter prediction method based on geometric geographic information. It comprehensively considers the intrinsic relationship between channel features and scattering environment features, integrates scene information into the network training process, and uses a dual-drive method of scene and data to achieve accurate and effective prediction of channel features in wireless communication scenarios.

[0076] 2. To achieve channel feature prediction based on geometric geographic information, this invention proposes an improved GAIN prediction network framework. The proposed network uses multi-channel feature data composed of scattering environment feature maps and partially sampled channel features as input, and transforms the prediction of channel feature parameters at different locations into image pixel values ​​for prediction. Attached Figure Description

[0077] Figure 1 This is a flowchart of the wireless channel feature parameter prediction method based on geometric geographic information proposed in this invention.

[0078] Figure 2 A schematic diagram illustrating the prediction of channel characteristic parameters for different height and vertical layers.

[0079] Figures 3a-3b This is a schematic diagram of the channel feature parameter prediction network framework proposed in this invention, wherein, Figure 3a For generator network structure, Figure 3b This is the discriminator network structure. Detailed Implementation

[0080] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings.

[0081] In one embodiment, such as Figure 1 The present invention proposes a method for predicting wireless channel feature parameters based on geometric geographic information, comprising the following steps:

[0082] Step 1: The user inputs initial parameters, including scene parameters and network parameters. Scene parameters include the length L, width W, and height H of the test region, the height layer grid resolution r1, the vertical layer grid resolution r2, and the sampling point position P. Network parameters include the number of training iterations i for the height and vertical layers, the error threshold ε, the maximum number of iterations l not exceeding the error threshold, and the generator learning rate α. GDiscriminator learning rate α D .

[0083] Step 2: Reconstruct the scene of the area to be measured according to user parameter settings. Horizontally layer the measurement data at different heights, and divide each height into grids according to a height-layer grid resolution r1, obtaining an M×N resolution scatterer occlusion mask matrix M and a scattering environment feature map E corresponding to each height h. The scatterer occlusion mask matrix M uses 1 and 0 to represent areas covered and uncovered by buildings, respectively. The scattering environment feature map E includes a feature map E describing the scatterer height. h Feature map E of scatterer type and location c Scatterer height characteristic map E h The height values ​​in the graph will be restricted to the range [0, 1] through normalization, and the calculation method is as follows:

[0084]

[0085] Where (i, j) represents the coordinates of the pixel. Let be the original height and the normalized height at point (i, j), respectively.

[0086] Scatterer type characteristic diagram E c The three types of scatterers—ground, buildings, and vegetation—are represented by a 0-2 classification label. The feature maps of the scatterers containing the three types are then encoded one-hot on the image pixels. This means that an image channel containing only that type is created for each type, and the image of each channel consists of 0 and 1, where 1 represents the location distribution of one type on the image.

[0087] Step 3: Based on the user-input sampling point location P, obtain the corresponding sampling matrix S at each height layer, and measure the channel impulse response of the receiver at the corresponding location. By extracting the channel feature parameters contained in the channel impulse response, construct a channel parameter feature map. Where K = 1, 2, 3 correspond to path loss, root mean square delay spread, and root mean square angle spread, respectively.

[0088] Step 4: Based on the sampled channel feature parameters, construct a Generative Adversarial Imputation Network (GAIN) to predict the channel feature parameters for the entire scene. First, predict the channel features at different height levels, and then combine the obtained scattering environment feature map E with the channel feature map including the partially sampled data. The combined feature data from multiple channels are used as conditional inputs to the improved GAIN network, and appropriate network parameters are set for training.

[0089] The specific steps for the fourth step are as follows:

[0090] S4.1: Randomly initialize the generator parameters θ of the network G and discriminator parameters θ D .

[0091] S4.2: Generator G network training and optimization. The generator G uses the input scattering environment features and a partially sampled channel feature map as conditional inputs to the network. Through training, it generates a channel feature image with all pixels padded. The calculation method is as follows:

[0092]

[0093] Where ⊙ represents the Hadamard product, the generator's output image is multiplied by M to obtain the channel feature prediction map X at the effective location outside the building. g The loss function of generator G consists of two parts. One part is the reconstruction error of the generated pixel values ​​at the original sampling location, which is expressed as the mean square error loss function between the generated value and the true value. The other part is the loss value of the discriminator's output for the generated pixel values, that is, the probability that the discriminator determines whether the channel characteristic at the sampling point location is the true value. The calculation method is as follows:

[0094]

[0095] Among them, D(·) ij Let represent the (i, j)th element of the decision matrix P output by the discriminator D, and λ represent the weights used to balance the two losses. The generator in the network should make it as indistinguishable as possible to the discriminator from real and generated pixel values. Considering that the discriminator has strong discriminative ability in the early stages of training of the generative adversarial network, and the generator has not yet learned the real distribution, log(1-D(X)) g ⊙(1-S)) ij The loss term will equal 0, and the network parameters of the generator and discriminator cannot be updated for this loss term. To avoid this situation, the first term of the generator loss function needs to be rewritten to minimize log(1-D(X)). g ⊙(1-S)) ij Change to maximize log(D(X)) g ⊙(1-S)) ij This means maximizing the probability that the generated pixel value is real, calculated as follows:

[0096]

[0097] in, Let m represent the loss function of the generator. ijLet M represent the (i, j)th element. The generator G uses a stochastic gradient optimizer to update the network parameters to minimize the negative value of the objective function. The calculation method is as follows:

[0098]

[0099] in, It is the updated generator parameter, α G It is the generator learning rate. This represents the negative gradient value of the generator G loss function.

[0100] S4.3: Discriminator D network training optimization. The input to discriminator D is the sampled channel feature parameters input by the user. The calculation method, combining the unsampled positions completed by the generator, is as follows:

[0101]

[0102] The discriminant matrix P output by discriminator D is calculated using the binary cross-entropy loss of the matrix elements corresponding to valid locations outside the building. The goal is to determine whether each padded pixel is a sampled or unsampled point as accurately as possible. The loss function of discriminator D is... The calculation method is as follows:

[0103]

[0104] Among them, s ij Let S represent the (i, j)th element. The discriminator D uses a stochastic gradient optimizer to update the network parameters to minimize the negative value of the objective function. The calculation method is as follows:

[0105]

[0106] in, These are the updated discriminator parameters, α D It is the discriminator learning rate. This represents the negative gradient value of the loss function of the discriminator D.

[0107] S4.4: Repeat the training process of S4.2 and S4.3, iterating until the network training count i is reached or the network loss function value does not exceed the threshold ε for l consecutive iterations, then the network training ends. This network is trained at different receiver heights, and the generator can output channel parameter prediction results for all locations at the same height layer based on the sampled channel feature images.

[0108] Step 5: The channel feature parameters sampled in Step 3 are vertically layered along the xOz plane perpendicular to the ground at a set resolution r², resulting in channel feature maps for different vertical layers. And the corresponding sampling matrix S′, and use the prediction network structure from step four to predict other non-sampling points in different vertical layers. The specific implementation steps of step five are as follows:

[0109] S5.1: Based on the reconstructed test area scene from the second step, obtain the vertical P×M resolution scatterer occlusion mask matrix M′ and scattering environment feature map E′, and then... M′ and E′ serve as conditional inputs to the improved GAIN prediction network.

[0110] S5.2: Randomly initialize the generator parameters θ of the network G′ and discriminator parameters θ D′ .

[0111] S5.3: The generator G′ and discriminator D′ networks are trained and optimized using the methods in S4.2 and S4.3.

[0112] S5.4: Repeat the training process in S5.3, iterating until the network training count i is reached or the network loss function value does not exceed the threshold ε for l consecutive iterations, then the network training ends. This network is trained on different vertical layers, and the generator can output channel parameter prediction results for all positions in different vertical layers based on the sampled channel feature images.

[0113] Step 6: Based on the channel feature prediction networks obtained in Steps 4 and 5 for different height and vertical layers, when the user inputs the location coordinates to be predicted in the scene, find the height or vertical layer where the location coordinates are located, and obtain the channel feature parameter prediction results for the corresponding location.

[0114] Next, this embodiment provides a specific example. In the scenario under test, there is one transmitter, and the transmitter's transmission frequency is set to f = 2.53 GHz.

[0115] Step 1: User inputs initial parameters. Scene parameters include the length L = 300m, width W = 250m, and height H = 100m of the area to be measured. The height layer mesh resolution is r1 = 2.5m, the vertical layer mesh resolution is r2 = 2.5m, and the sampling point location P of the scene to be measured is as follows: Figure 2 As shown; network parameters include height layer and vertical layer; network training times i = 3000; error threshold ε = 0.001; maximum number of iterations not exceeding the error threshold l = 20; generator learning rate α. G =0.0001, discriminator learning rate α D =0.0001.

[0116] Step 2: Scene reconstruction is performed on the test area. The test area at different height levels is divided into meshes with a resolution of 2.5m, resulting in a scatterer occlusion mask matrix M and a scattering environment feature map E of size 120×100 for each height level. The normalized scatterer height feature map E is then calculated. h Furthermore, the feature maps classified according to the scatterer type are encoded using one-hot encoding, and a feature map E of the scatterer type is constructed by creating an image channel with only that type for each category.

[0117] Step 3: Based on the channel characteristic parameters collected at position P, construct a channel parameter feature map including path loss, root mean square delay spread, and root mean square arrival horizontal angle spread.

[0118] Step 4: Construct an improved GAIN network structure that combines scattering environment feature maps at different height levels with channel feature maps including partial samples. The network is trained by combining the conditional inputs of the combined GAIN network to predict channel parameters at unsampled locations. The network layers are configured as follows: Figure 3a and Figure 3b As shown, k represents the kernel size of the convolutional layer, n represents the number of channels, and s represents the convolution stride. The specific implementation steps are as follows:

[0119] S4.1: Randomly initialize the generator G parameters θ of the network G and discriminator D parameter θ D .

[0120] S4.2: Combine the scattering environment feature map from step 2 and the sampling channel feature map from step 3 to form multi-channel feature data as input to generator G, and minimize the loss function of generator G. And update the generator network parameters θ G .

[0121] S4.3: Sample the channel feature map from step three. The channel feature map, which is the sampled portion of the output from generator S4.2, is combined with the input of discriminator D to minimize the loss function of discriminator D. And update the discriminator D network parameters θ G .

[0122] S4.4: Repeat the training process of S4.2 and S4.3, iterating until the network training times reach i = 3000 or the network loss function value does not exceed the threshold ε = 0.001 after l = 20 consecutive iterations. Then, end the network training and obtain the network output prediction result as follows. Figure 2 As shown.

[0123] Step 5: The channel feature parameters sampled in Step 3 are layered along the xOz plane perpendicular to the ground at a set height resolution of 2.5m, and the channel feature map of this vertical plane is obtained. Given the corresponding sampling matrix S′, the prediction network structure from step four is used to predict other non-sampling points in the vertical plane. The specific implementation steps are as follows:

[0124] S5.1: Simultaneously, based on the reconstructed test area scene from the second step, a vertically 40×120 resolution scatterer occlusion mask matrix M′ and a scattering environment feature map E′ are obtained. M′ and E′ serve as conditional inputs to the improved GAIN prediction network.

[0125] S5.2: Randomly initialize the generator parameters θ′ of the network G and discriminator parameters θ′ D .

[0126] S5.3: The generator G′ and discriminator D′ networks are trained and optimized using the methods in S4.2 and S4.3.

[0127] S5.4: Repeat the training process in S5.3, iterating until the network training count i = 5000 or the network loss function value does not exceed the threshold ε = 0.001 after 20 consecutive iterations. The network is then trained on different vertical layers, and the generator can output channel parameter prediction results for all positions in different vertical layers based on the sampled channel feature images.

[0128] Step 6: Based on the channel feature prediction networks obtained in Step 4 and Step 5 for different height and vertical layers, when the user inputs the location coordinates to be predicted in the scene, the height or vertical layer where the location coordinates are located is found, and the channel feature parameter prediction results at the corresponding location are obtained. The prediction results for the height layer are shown in Table 1, and the prediction results for the vertical layer are shown in Table 2.

[0129] Table 1 Prediction results for the upper-level section

[0130]

[0131] Table 2 Prediction results for vertical layers

[0132]

[0133] In another embodiment, the present invention provides a computer-readable storage medium storing a computer program that causes a computer to execute the wireless channel characteristic parameter prediction method based on geometric geographic information as described in Embodiment 1.

[0134] In another embodiment, the present invention proposes an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the wireless channel feature parameter prediction method based on geometric geographic information as described in Embodiment 1.

[0135] In the embodiments disclosed in this application, a computer storage medium may be a tangible medium that may contain or store programs for use by or in conjunction with an instruction execution system, apparatus, or device. The computer storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of computer storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0136] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed in this application can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0137] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principle of the present invention should be considered within the scope of protection of the present invention.

Claims

1. A method for predicting wireless channel characteristic parameters based on geometric geographic information, characterized in that, include: Step 1: Input initial parameters, including scene parameters and network parameters; Step 2: Reconstruct the scene of the area to be tested based on the input initial parameters; Step 3: In the reconstructed test area scene, based on the input sampling point location, the channel impulse response at different height layers is measured to obtain the corresponding location, and the channel feature parameters contained in the channel impulse response are extracted; specifically as follows: Based on the location of the sampling points, a corresponding sampling matrix is ​​obtained for each height layer. The channel impulse response at the corresponding receiver location is obtained by measurement, and a channel parameter feature map is constructed by extracting the channel feature parameters contained in the channel impulse response. ,in These correspond to path loss, root mean square delay spread, and root mean square angle spread, respectively. Step 4: Based on the channel feature parameters sampled in Step 3, construct a generative adversarial completion network for predicting channel feature parameters at different height levels; Step 5: Vertically layer the channel feature parameters sampled in Step 3 along the direction perpendicular to the ground to obtain channel feature parameters for different vertical layers. Based on the channel feature parameters of different vertical layers, construct a generative adversarial completion network for predicting the channel feature parameters of different vertical layers; specifically including the following steps: S5.1: Channel parameter feature map of the third sampling step According to vertical layer grid resolution along the direction perpendicular to the ground Further vertical layering was performed, and channel feature maps of different vertical layers were obtained. and the corresponding sampling matrix ,in These correspond to path loss, root mean square delay spread, and root mean square angle spread, respectively. Combined with the reconstructed test area scene from the second step, the vertical [value] is obtained. Scattering occlusion mask matrix with resolution and scattering environment feature map ; S5.2: Randomly initialize the generator parameters of the network and discriminator parameters ; S5.3: For generators and discriminator Network training optimization: 1) Generator The network training and optimization process is as follows: generator Based on the scattering environment characteristic map and channel parameter feature maps including partial sampling As input to the network, a channel feature image with all pixels padded is generated during training. The calculation method is as follows: in, The output image of the generator represents the Hadamard product. and Multiply to obtain the channel feature prediction map at the effective location outside the building. ; generator loss function as follows: in, express The One element, Discriminator Output decision matrix The One element, This represents the weight values ​​used to balance the two parts of the loss; generator The stochastic gradient optimizer is used to update the parameters in order to minimize the negative value of the objective function. The calculation method is as follows: in, These are updated generator parameters. It is the generator learning rate. Represents generator The negative gradient value of the loss function; 2) Discriminator The network training and optimization process is as follows: Discriminator Input It is a channel parameter feature map including partial sampling. and through generator The combination of the completed unsampled locations is calculated as follows: Discriminator Output discriminant matrix It is to calculate the binary cross-entropy loss for the matrix elements corresponding to the effective locations outside the building; Discriminator loss function as follows: in, express The One element; Discriminator The stochastic gradient optimizer is used to update the parameters in order to minimize the negative value of the objective function. The calculation method is as follows: in, These are the updated discriminator parameters. It is the discriminator learning rate. Discriminator The negative gradient value of the loss function; S5.4: Repeat the training process of S5.3 until the iteration condition is met to obtain a generative adversarial completion network for predicting channel feature parameters of different vertical layers; Step 6: Based on the generative adversarial completion networks with different height layers and different vertical layers constructed in Steps 4 and 5, predict the channel feature parameters of the location coordinates to be predicted in the scene.

2. The wireless channel characteristic parameter prediction method based on geometric geographic information as described in claim 1, characterized in that: In the first step, the scene parameters include the length, width, height, height layer grid resolution, vertical layer grid resolution, and sampling point location of the region to be tested; the network parameters include the number of network training iterations for the height layer and vertical layer, the error threshold, the maximum number of iterations not exceeding the error threshold, the generator learning rate, and the discriminator learning rate.

3. The wireless channel characteristic parameter prediction method based on geometric geographic information as described in claim 1, characterized in that: In the second step, the scene reconstruction of the area to be tested is performed as follows: The channel impulse response measured at different altitudes is horizontally layered, and each altitude layer is processed according to the altitude layer grid resolution. Perform mesh generation to obtain each height layer. corresponding Scattering occlusion mask matrix with resolution and scattering environment feature map Among them, the scatterer occlusion mask matrix Using 1 and 0 to represent areas covered and uncovered by buildings, respectively, a scattering environment feature map is shown. Includes a scatterer height feature map used to describe the scatterer height. and scatterer type feature map used to describe the location of scatterer types .

4. The wireless channel characteristic parameter prediction method based on geometric geographic information as described in claim 3, characterized in that: The scatterer height feature map By normalizing, the height values ​​in the graph are limited to a certain range. Within the specified range, the calculation method is as follows: in, Represents the coordinates of a pixel. Points The original height and normalized height at the location; The scatterer type characteristic diagram The three types of scatterers—ground, buildings, and vegetation—are represented by a classification label of 0-2. An image channel containing only that type is created for each type. The image in each channel consists of 0s and 1s, where 1s represent the location distribution of the type in the image.

5. The wireless channel characteristic parameter prediction method based on geometric geographic information as described in claim 1, characterized in that: The fourth step specifically includes the following steps: S4.1: Randomly initialize the generator parameters of the network and discriminator parameters ; S4.2: For the generator The network training and optimization process is as follows: generator Based on the scattering environment characteristic map and channel parameter feature maps including partial sampling As input, a channel feature image with all pixels padded is generated through training, and the calculation method is as follows: in, The output image of the generator represents the Hadamard product. and Multiply to obtain the channel feature prediction map at the effective location outside the building. ; generator loss function as follows: in, express The One element, Discriminator D Output decision matrix The One element, This represents the weight values ​​used to balance the two parts of the loss; generator The stochastic gradient optimizer is used to update the parameters in order to minimize the negative value of the objective function. The calculation method is as follows: in, These are updated generator parameters. It is the generator learning rate. Represents generator The negative gradient value of the loss function; S4.3: Discriminator The network training and optimization process is as follows: Discriminator Input It is a channel parameter feature map including partial sampling. and through generator The combination of the completed unsampled locations is calculated as follows: Discriminator Output discriminant matrix It is to calculate the binary cross-entropy loss for the matrix elements corresponding to the effective locations outside the building; Discriminator loss function as follows: in, express The One element; Discriminator The stochastic gradient optimizer is used to update the parameters in order to minimize the negative value of the objective function. The calculation method is as follows: in, These are the updated discriminator parameters. It is the discriminator learning rate. Discriminator The negative gradient value of the loss function; S4.4: Repeat the training process of S4.2 and S4.3 until the iteration condition is met, then the network training ends and a generative adversarial completion network is obtained for predicting channel feature parameters at different height levels.

6. The wireless channel characteristic parameter prediction method based on geometric geographic information as described in claim 1, characterized in that: The iteration condition is the number of network training iterations. Reaching the set number of times, or consecutively The network loss function value in the next iteration does not exceed the error threshold. , The maximum number of iterations that does not exceed the error threshold.

7. A computer-readable storage medium storing a computer program, characterized in that, The computer program causes the computer to execute the wireless channel characteristic parameter prediction method based on geometric geographic information as described in any one of claims 1-6.

8. An electronic device, characterized in that, include: The memory, the processor, and the computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the wireless channel characteristic parameter prediction method based on geometric geographic information as described in any one of claims 1-6.

Images