A base station optimization site selection method and system based on transformer network

Through the optimization site selection method of base stations based on Transformer network, the characteristics of building distribution maps are extracted and the candidate area dependencies are learned, and the problem of low site selection efficiency in the existing technology is solved, and efficient and accurate base station location recommendation and radio map are achieved. Generate, significantly shortening the time for base station deployment tasks.

CN118488465BActive Publication Date: 2024-12-03JIANGHAN UNIVERSITY
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202410239964.5
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-03-04
Publication Date
2024-12-03
Estimated Expiration
2044-03-04

AI Technical Summary

Technical Problem

The existing base station site selection method based on convolutional neural network is inefficient in downstream task applications, requiring multiple iterations to obtain the optimal base station location, resulting in wasted computing and time resources, and it is impossible to effectively unify radio map estimation and base station deployment tasks. .

Method used

Adopting the optimization site selection method of base stations based on Transformer network, by obtaining the building distribution map, extracting features and generating image embedding vectors, learning the dependencies of candidate areas, outputting radio feature vectors and base station location encoding, and using the self-attention mechanism to correlate the global environment Features, recommend the optimal base station coordinates and generate high-definition radio maps.

Benefits of technology

It greatly improves the efficiency and accuracy of base station site selection, reduces calculation overhead, realizes the accuracy of radio map estimation and maximizes signal coverage, and significantly shortens the running time of base station deployment tasks in urban areas.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118488465B_ABST
    Figure CN118488465B_ABST
Patent Text Reader

Abstract

The present invention provides a method and system for optimizing the site selection of base stations based on a transformer network. The method comprises the following steps: a backbone network extracts the features of a building distribution map and inputs them into an image feature encoder, outputs an image embedding vector of the building distribution map to an image decoder; a candidate area is input into a prompt encoder, and a region embedding vector of the candidate area is output to an image decoder; the image decoder arranges the image embedding vector and the region embedding vector in order, and outputs a radio feature vector and a base station location code; a first feedforward network transforms the base station location code dimension and outputs it to a recommendation module; the recommendation module selects the optimal base station coordinates and their corresponding radio feature vectors; a second feedforward network generates a weight matrix; and a feature pyramid module generates a radio map corresponding to the optimal base station. The present invention can improve the efficiency and accuracy of base station site selection calculations while ensuring the accuracy of radio map estimation.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of radio technology, and in particular relates to a base station optimization site selection method and system based on a transformer network. Background Art

[0002] In the context of the rapidly developing digital age, wireless communication technology is in a period of great transformation. With the acceleration of urbanization, we are facing an increasingly complex urban environment, which poses unprecedented challenges to the design and maintenance of wireless networks. In this environment, effective network planning and signal coverage estimation have become key tasks. The wireless propagation model is a mathematical model used to describe the propagation process of radio waves in space. It can quantify the attenuation, dispersion and reflection of radio waves from the transmitter to the receiver, and is widely used in network planning and design, communication performance evaluation, fault diagnosis and optimal deployment and other tasks.

[0003] In recent years, data-driven deep learning frameworks have been widely used in radio map estimation. For example, the underlying laws of radio propagation are extracted through convolutional neural networks. It can directly generate a radio map of the entire area through base station locations and building maps, and has achieved state-of-the-art performance on public datasets. On this basis, the building map and radio map are simultaneously generated through RSRP measurements, and sampling values ​​are introduced to correct the radio map. Alternatively, the high-frequency details of the radio map are restored by introducing a generative adversarial network (GAN), and the impact of uneven RSRP sampling areas on the shadow effect is explored.

[0004] However, in the application process of downstream tasks, such as base station deployment, these CNN-based methods are limited by the model input, that is, the initial base station location and the electronic environment map must be input simultaneously for fitting. Due to the variable location of base stations, the efficiency of sequentially searching for the optimal base station is very low. Because of this, they often require multiple iterations to obtain the optimal base station location, resulting in a large waste of computing and time resources, and limiting the widespread application of radio maps. Therefore, it is urgent to explore an end-to-end framework that can accept multi-modal data input and unify wireless propagation models and downstream tasks. Summary of the invention

[0005] The purpose of the present invention is to solve the deficiencies of the above-mentioned background technology and to provide a base station optimization site selection method and system based on transformer network, which can improve the efficiency and accuracy of base station site selection calculation while ensuring the accuracy of radio map estimation.

[0006] The technical solution adopted by the present invention is: a base station optimization site selection method based on transformer network, comprising the following steps:

[0007] Obtain a building distribution map of the expected radiation area of ​​the base station to be deployed;

[0008] Annotate candidate areas of base stations using bounding boxes in the building distribution map;

[0009] The backbone network extracts the features of the building distribution map and inputs them into the image feature encoder, and outputs the image embedding vector of the building distribution map to the image decoder;

[0010] Input the candidate region to the hint encoder, and output the region embedding vector of the candidate region to the image decoder;

[0011] The image decoder arranges the image embedding vector and the region embedding vector in order, learns the dependency between them, and outputs the radio feature vector and the base station location code.

[0012] The first feedforward network transforms the base station location encoding dimension and outputs it to the recommendation module;

[0013] The recommendation module takes the maximization of available signal coverage as the standard and outputs the optimal base station coordinates and their corresponding radio feature vectors based on the base station location encoding after dimension transformation;

[0014] The second feed-forward network generates a weight matrix based on the base station position code;

[0015] The feature pyramid module generates a radio map corresponding to the optimal base station based on the radio feature vector corresponding to the optimal base station, the weight matrix and the features of the building distribution map.

[0016] In the above technical solution, the training process of the transformer network model includes: first freezing the recommendation module and the first feedforward network, training the backbone network, image feature encoder, prompt encoder, image decoder, second feedforward network and feature pyramid module, so that the transformer network model has the ability to generate high-definition radio maps; then freezing the backbone network, image feature encoder, prompt encoder, image decoder, second feedforward network and feature pyramid module, training the recommendation module and the first feedforward network, so that the transformer network model has the ability to perceive the optimal base station.

[0017] In the above technical solution, the process of training the backbone network, the image feature encoder, the hint encoder, the image decoder, the second feedforward network and the feature pyramid module includes:

[0018] Constructing a first training set; a single sample of the first training set includes a building distribution map and base station coordinates as model inputs, and a corresponding real radio map as a training label;

[0019] The backbone network, the image feature encoder, the cue encoder, the image decoder, the second feedforward network and the feature pyramid module are trained using the first training set; the feature pyramid module outputs a reconstructed radio map;

[0020] After the loss function is calculated by comparing the reconstructed radio map with the true radio map, the gradient updates the parameters of the backbone network, image feature encoder, hint encoder, image decoder, second feedforward network and feature pyramid module.

[0021] In the above technical solution, the process of training the recommendation module includes:

[0022] Construct the second training set: randomly select bounding boxes in the building distribution map, and obtain several candidate base station sites by sampling at equal intervals; calculate the radio map corresponding to each base station site; based on the radio map, calculate the ranking matrix corresponding to each base station site with the maximization of the available signal coverage as the standard; use the building distribution map and the candidate base station sites obtained based on the bounding box as model inputs, and use the ranking matrix as training labels;

[0023] Using the second training set to train the recommendation module and the first feed-forward network;

[0024] The image decoder outputs base station position codes corresponding to a number of candidate base station sites;

[0025] The base station location code is transformed into a score matrix through the first feed-forward network and then passed through the recommendation module;

[0026] After the loss function is calculated for the score matrix and the ranking matrix, the parameters of the recommendation module and the first feedforward network are updated with gradients.

[0027] In the above technical solution, the backbone network initially extracts the features of the building distribution map and then outputs a set of image feature matrices, and obtains a set of image embedding vectors based on the image feature division, each of which represents a sub-region mapped to the original image space; the image feature decoder is based on the self-attention mechanism and encodes the set of image embedding vectors through the visual transformer to reduce image redundancy.

[0028] In the above technical solution, the encoder is prompted to evenly scatter points within the bounding box with a set step size to obtain several candidate base station positions; then the position coding vectors of several candidate base stations are fused and superimposed with the position coding of the upper left corner and the lower right corner of the bounding box to obtain the regional embedding vector representing the positions of several candidate base stations.

[0029] In the above technical solution, the image decoder calculates the importance weights between the feature vector of each image embedding vector and the feature vector of each region embedding vector through an attention mechanism; based on all paired importance weights and a nonlinear ReLU layer, through a multi-layer attention mechanism, the feature vectors of the image embedding vector and the region embedding vector are updated to radio feature vectors and base station location codes.

[0030] In the above technical solution, the recommendation module obtains the score matrix corresponding to all candidate base stations under the available signal coverage range standard based on the base station position code after the dimension transformation; and obtains the radio feature vector corresponding to the optimal base station through the optimal selection operation.

[0031] In the above technical solution, the feature pyramid module downsamples the features of the building distribution map; restores the radio feature vector corresponding to the optimal base station into image features; and upsamples the image features layer by layer:

[0032] The downsampled features of the building distribution map at the corresponding level are superimposed and fused with the image features, and then upsampled. The upsampled features are added and fused with the information of the channel dimension supplemented by pixel shuffle as new image features.

[0033] The above upsampling process is performed repeatedly until the image features of the original building distribution map size are obtained;

[0034] The image features of the original building distribution map size are multiplied by the weight matrix to obtain the radio map corresponding to the optimal base station.

[0035] The present invention also provides a base station optimization site selection system based on a transformer network, comprising a backbone network, an image feature encoder, a prompt encoder, an image decoder, a first feedforward network, a second feedforward network, a recommendation module and a feature pyramid module;

[0036] The backbone network is used to extract the features of the building distribution map and input them into the image feature encoder.

[0037] The image feature encoder is used to output the image embedding vector of the building distribution map to the image decoder;

[0038] The hint encoder is used for outputting a region embedding vector of the candidate region to the image decoder according to the candidate region;

[0039] The image decoder is used to arrange the image embedding vector and the region embedding vector in order, learn the dependency between the two, and output the radio feature vector and the base station location code;

[0040] The first feed-forward network is used to transform the base station location encoding dimension and then output it to the recommendation module;

[0041] The recommendation module is used to select the optimal base station coordinates and their corresponding radio feature vectors based on the base station location codes after dimension transformation, taking the maximization of the available signal coverage as the standard;

[0042] The second feed-forward network is used to generate a weight matrix based on the base station position code;

[0043] The feature pyramid module is used to generate a radio map corresponding to the optimal base station based on the radio feature vector corresponding to the optimal base station, the weight matrix and the features of the building distribution map.

[0044] The beneficial effects of the present invention are as follows: the present invention is based on the transformer network, and by prompting the encoder to encode all candidate base station locations in the entire area, all high-definition radio maps in the area can be obtained through only one forward reasoning, which greatly reduces the computational overhead. Benefiting from the self-attention mechanism of the transformer structure, base station point selection can be associated with global environmental features without considering their distance in the sequence, which greatly increases the accuracy of radio map estimation. The end-to-end deep learning model designed for the base station deployment task of the present invention effectively unifies the radio map estimation process with the downstream base station deployment task. It can quickly recommend the best point from a given candidate base station to maximize signal coverage. For a fixed number of candidate points in an urban cell, the present invention allows the user to define a base station candidate box of any size and any position. The present invention can quickly recommend a predefined number of base station locations in the candidate box to maximize the signal coverage. By reducing computational overhead, it can help shorten the running time of base station deployment tasks in urban areas.

[0045] Furthermore, the present invention first trains the teacher model and then trains the student model. The purpose of training the teacher model is to obtain a high-precision radio map estimation model that can obtain accurate radio maps through building maps and the locations of base stations to be deployed. The purpose of training the student model is to enable the model to have the ability to perceive the optimal base station. The sequential training method adopted by the present invention couples the radio estimation model and the base station deployment task, significantly shortening the running time of the base station deployment task in urban areas.

[0046] Furthermore, the present invention trains the teacher model, using the RadioMapSeer public data set to provide the correspondence between building maps, radio maps and base station locations, which can effectively guide the training of the model. During the training process, the present invention freezes the base station scoring network, trains the backbone network, image encoder, prompt encoder, image decoder and feature pyramid module of the model, so that the model has the ability to reconstruct the radio map.

[0047] Furthermore, the present invention trains the student model and uses the RadioMapSeer public data set to guide the model training. During the training process, the present invention first creates a copy of the teacher model and freezes the weights of the teacher model to make the training process more stable. Secondly, the present invention restores the training gradient of the base station scoring network, freezes the relevant weights of the radio map reconstruction, and specifically trains the model to perceive the optimal base station, thereby ensuring the efficiency and effectiveness of model training. Finally, the present invention uses the base station score matrix obtained by cyclically running the teacher model as the true value label, and gradually approaches the teacher model's ability to deploy base stations through continuous iterative learning.

[0048] Furthermore, the backbone network and transformer-based image encoder used in the present invention can effectively extract the features of the building map and encode the environmental information into a high-dimensional vector, effectively helping the model to simulate the underlying physical laws of radio propagation. At the same time, benefiting from the self-attention mechanism of the transformer model, the model can associate global environmental features without considering their distance in the sequence, greatly increasing the accuracy of radio map estimation.

[0049] Furthermore, the present invention completes the process of generating regional embedding vectors based on the prompt encoder, maps the base station location to a high-dimensional space, can be associated with building features, and learns the propagation law of radio. It is worth noting that the prompt encoder not only embeds location information, but also constructs embedding vectors of multimodal information such as text information and image information, which greatly improves the application potential of the radio estimation model.

[0050] Furthermore, the image decoder of the present invention interactively models the building features and the regional embedding vector, and learns the dependency between the base station location, building information and the signal receiving point through the cross-attention mechanism, so as to ensure that the output base station site selection conclusion fully considers the characteristics of the building distribution and the base station. In addition, the present invention also generates a score matrix and calculates the importance weight through a feedforward network. Among them, the score matrix can select the best point among the candidate base stations and maximize the signal coverage of the base station, and the importance weight can better restore the details of the radio map.

[0051] Furthermore, the present invention reconstructs a high-resolution radio map through a feature pyramid module. The present invention first uses transposed convolution to establish a mapping relationship between image features and high-definition images, and uses jump connections to introduce shallow information to better utilize the fine-grained information of features. At the same time, the present invention also uses the PixelShuffle method to supplement the channel information lost in the upsampling process to achieve a more powerful feature representation. The present invention not only improves the accuracy of radio map reconstruction, but also improves the accuracy of base station site selection, and reduces the time required for base station site selection, greatly increasing the application potential of radio maps. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 It is a schematic diagram of the method flow of the present invention;

[0053] Figure 2 It is a schematic diagram of the system architecture of the present invention;

[0054] Figure 3 It is a schematic diagram of the training process of the present invention;

[0055] Figure 4 It is a schematic diagram of a feature pyramid module of the present invention;

[0056] Figure 5 A comparison chart of the training loss function of the present invention and other existing models;

[0057] Figure 6 A comparison diagram of dense mode reasoning between the present invention and other existing models;

[0058] Figure 7 A schematic diagram of a specific application implementation. DETAILED DESCRIPTION

[0059] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments to facilitate a clear understanding of the present invention, but they do not constitute a limitation on the present invention.

[0060] like Figure 1 As shown, the present invention provides a base station optimization site selection method based on a transformer network, comprising the following steps:

[0061] Obtain a building distribution map of the expected radiation area of ​​the base station to be deployed;

[0062] Annotate candidate areas of base stations using bounding boxes in the building distribution map;

[0063] The transformer network model includes a backbone network, an image feature encoder, a prompt encoder, an image decoder, a first feedforward network, a second feedforward network, a recommendation module, and a feature pyramid module;

[0064] The backbone network extracts the features of the building distribution map and inputs them into the image feature encoder, and outputs the image embedding vector of the building distribution map to the image decoder;

[0065] Input the candidate region to the hint encoder, and output the region embedding vector of the candidate region to the image decoder;

[0066] The image decoder arranges the image embedding vector and the region embedding vector in order, learns the dependency between them, and outputs the radio feature vector and the base station location code.

[0067] The first feedforward network transforms the base station location encoding dimension and outputs it to the recommendation module;

[0068] The recommendation module takes the maximization of available signal coverage as the standard and selects the optimal base station coordinates and their corresponding radio feature vectors based on the base station location codes after dimension transformation;

[0069] The second feed-forward network generates a weight matrix based on the base station position code;

[0070] The feature pyramid module generates a radio map corresponding to the optimal base station based on the radio feature vector corresponding to the optimal base station, the weight matrix and the features of the building distribution map.

[0071] like Figure 2 As shown, a base station optimization site selection system based on a transformer network includes a backbone network (ResNet), an image feature encoder (Transformer Encoder), a prompt encoder (Prompt Encoder), an image decoder (Transformer deccoder), a first feedforward network (FFN-1), a second feedforward network (FFN-1), a recommendation module and a feature pyramid module;

[0072] The backbone network is used to extract the features of the building distribution map and input them into the image feature encoder.

[0073] The image feature encoder is used to output the image embedding vector of the building distribution map to the image decoder;

[0074] The hint encoder is used for outputting a region embedding vector of the candidate region to the image decoder according to the candidate region;

[0075] The image decoder is used to arrange the image embedding vector and the region embedding vector in order, learn the dependency between the two, and output the radio feature vector and the base station location code;

[0076] The first feed-forward network is used to transform the base station location encoding dimension and then output it to the recommendation module;

[0077] The recommendation module is used to select the optimal base station coordinates and their corresponding radio feature vectors based on the base station location codes after dimension transformation, taking the maximization of the available signal coverage as the standard;

[0078] The second feed-forward network is used to generate a weight matrix based on the base station position code;

[0079] The feature pyramid module is used to generate a radio map corresponding to the optimal base station based on the radio feature vector corresponding to the optimal base station, the weight matrix and the features of the building distribution map.

[0080] The method proposed in the present invention is implemented based on the system proposed in the present invention. The present invention takes into account both the accuracy of radio map estimation and the computational and time complexity of determining the optimal base station location. The transformer network model only requires a global building map and candidate area boxes as input, avoiding the need for additional data. The present invention encodes the candidate areas of the base station into a set of high-dimensional vectors, and queries the optimal base station location coordinates in the high-dimensional features of the building distribution through a decoder. During the query process, this specific embodiment regards areas with signal strength exceeding -105dBm as available communication coverage areas, and selects the optimal base station location from the radio map based on this standard to maximize signal coverage.

[0081] Benefiting from the inherent attention mechanism, the transformer network model exceeds the kernel size limitation of the traditional convolutional neural network and allows global feature modeling without considering the sequence distance. At the same time, the transformer network model has a flexible prompt encoder that supports three different input modes: sparse (point-based), dense (box-based) and none. The present invention uses single-point input to complete the training of the transformer network model; using the input, the optimal radio map customized for the specified area is realized.

[0082] Specifically, given a global building distribution map As the input of the model; the backbone network can use ResNet34. The backbone network initially extracts the building distribution map features and outputs a set of image feature matrices Among them, H Z =16,W Z =16, C = 256. According to the image feature Z, a set of image embedding vectors Ι = {z1, z2, z3, ..., z h×w}, each image embedding vector Represents a sub-region mapped to the original image space. The image feature decoder is based on the self-attention mechanism. It encodes the image embedding vector of this group through the visual transformer pre-trained by MAE to obtain Ι′={z′1,z′2,z′3,...,z′ h×w} to reduce image redundancy. In this process, benefiting from the self-attention mechanism of the Transformer structure, the Visual Transformer (ViT) is free from the limitation of the convolution kernel size and can encode global features without considering their distance in the sequence.

[0083] Specifically, during the use of the prompt encoder, the operator selects a bounding box and inputs an instruction to determine the base station selection accuracy, that is, to select the best base station from several candidate base stations in the bounding box. Based on the input instruction, the prompt encoder evenly scatters points in the bounding box with a set step size to obtain several candidate base station positions; then the position encoding vectors of several candidate base stations are The position encoding of the upper left corner of the bounding box and the position code of the lower right corner Fusion and superposition are performed to obtain the regional embedding vector A = {a1, a2, ..., a N+1 ,a N+2}.

[0084] Specifically, the image decoder calculates the importance weights between the feature vector of each image embedding vector and the feature vector of each region embedding vector through an attention mechanism; based on all pairwise importance weights and a nonlinear ReLU layer, through a multi-layer attention mechanism, the feature vectors of the image embedding vector and the region embedding vector are updated to the radio feature vector and the base station location code.

[0085] Assume that the eigenvector of the image embedding vector Ι′={z′1,z′2,z′3,...,z′ h×w} and the feature vector A = {a1, a2, o} of the region embedding vector as the input of the image decoder, where Represents the output token. In the image decoder, it consists of L layers of cross-attention mechanisms. The feature vector h is calculated through the attention mechanism i ∈Ι′ and h j The importance weight α between ∈A ij . When all pairs of α ij After calculation, use α ij And the nonlinear ReLU layer, all h i Update to h i ′, the calculation formula is as follows:

[0086]

[0087]

[0088]

[0089] Among them, W k ,W q ,W v They are the key, query and value weight matrices, W r and W o is the transformation matrix, b1 and b2 are the deviation matrices. The above learnable matrices This is the key to the attention mechanism. h×w ,a1,a2,o} as the input of the next layer, and after repeating L times, the final output H′={z″1,...,z″ h×w ,a′1,a′2,o′}.

[0090] Specifically, the recommendation module is based on the base station location encoding after dimension transformation Get the score matrix corresponding to all candidate base stations under the available signal coverage standard The radio feature vector corresponding to the optimal base station is obtained by the optimal selection operation Ι″={z″1,z″2,z″3,...,z″ h×w}.

[0091] Specifically, the function of the feature pyramid module is to convert the radio feature vector corresponding to the optimal base station into {z″1, z″2, z″3, ..., z″ h×w}Reconstructed into image features That is, the radio map. In order to better learn the representation of features at different scales, such as Figure 3 As shown in the figure, the feature pyramid module downsamples the features of the building distribution map; for image features Upsampling layer by layer:

[0092] The downsampled features Z of the building distribution map at the corresponding level are superimposed and fused with the image features F by skip connection, and then upsampled. The mapping relationship from low-resolution features to high-resolution features is established by transposed convolution. The upsampled features are added and fused with the information of the channel dimension supplemented by pixel shuffle as new image features, thereby achieving more powerful feature representation.

[0093] The above upsampling process is repeated until the image features of the original building distribution map size are obtained.

[0094] The second feedforward network FFN2 generates a weight matrix based on the base station location code It can represent the feature matrix F out The importance weight in the channel dimension is assigned by transforming the dimension of the position encoding;

[0095] The image features of the original building distribution map size are multiplied by the weight matrix to obtain the radio map corresponding to the optimal base station.

[0096] like Figure 4 As shown, the training process of the transformer network model includes:

[0097] In training step 1, the radio map is pre-trained: first, the recommendation module and the first feed-forward network FFN1 are frozen, and the backbone network, image feature encoder, prompt encoder, image decoder, second feed-forward network FFN2 and feature pyramid module are trained to enable the transformer network model to generate high-definition radio maps;

[0098] In training step 2, training is performed to recommend the best base station location from the given N candidate base stations: freezing the backbone network, image feature encoder, prompt encoder, image decoder, second feedforward network FFN2 and feature pyramid module, training the recommendation module and the first feedforward network FFN1, so that the transformer network model has the ability to perceive the optimal base station.

[0099] In training step 1, the process of training the backbone network ResNet, image feature encoder, hint encoder, image decoder, second feedforward network and feature pyramid module includes:

[0100] Constructing a first training set; a single sample of the first training set includes a building distribution map and base station coordinates as model inputs, and a corresponding real radio map as a training label; the present invention uses real open source data to construct the first training set;

[0101] The first training set is used to train the backbone network, the image feature encoder, the prompt encoder, the image decoder, the second feedforward network and the feature pyramid module;

[0102] The backbone network extracts the features of the building distribution map and inputs them into the image feature encoder, which outputs the image embedding vector of the building distribution map to the image decoder; the prompt encoder outputs the region embedding vector to the image decoder according to the base station coordinates;

[0103] The image decoder arranges the image embedding vector and the region embedding vector in order, learns the dependency between the two, and outputs the radio feature vector

[0104] The second feed-forward network FFN2 generates a weight matrix based on a single base station position code;

[0105] Since there is only one base station at this time, the feature pyramid module is based on the feature Z of the building distribution map and the radio feature vector and the weight matrix outputs the reconstructed radio map;

[0106] After calculating the SSIM loss function between the reconstructed radio map and the true radio map, the gradient updates the parameters of the backbone network, image feature encoder, hint encoder, image decoder, second feedforward network and feature pyramid module.

[0107] Since radio map estimation is a regression fitting problem for each pixel in the building distribution map, this specific embodiment uses SSIM as the loss function to consider the convergence of the model from three perspectives: brightness, contrast, and structure. Its calculation formula is as follows:

[0108] l ssim (p,g)=l(p,g)×c(p,g)×s(p,g)

[0109] Where p and g correspond to the pixel value estimated by the model and the true pixel value of the training label, respectively. l(p,g), c(p,g), and s(p,g) quantitatively calculate the similarity between the predicted value and the true value in terms of brightness, contrast, and structure, respectively.

[0110] The calculation formula of l(p,g) is as follows:

[0111]

[0112]

[0113] C1=(K1L) 2

[0114] in, pi Represents a single pixel value, μ p and μ g They represent the average of N pixel values ​​in the model prediction value and the true value of the training label respectively. C1 is a hyperparameter set to prevent the denominator from being zero. Set K1=0.01 and L=255.

[0115] c(p,g) is used to characterize the intensity of the image brightness change, that is, the standard deviation of the pixel value. Its calculation formula is as follows:

[0116]

[0117]

[0118] C2=(K2L) 2

[0119] Among them, σ p and σ gRepresents the standard deviation of N pixels in the model prediction value and the true value of the training label, respectively. C2 is a hyperparameter, and K2=0.03 and L=255 are set. s(p,g) uses the vector composed of all pixels in the image to calculate the cosine similarity between the predicted value and the true value. The calculation formula is as follows:

[0120]

[0121]

[0122] Among them, σ pg Represents the covariance of N pixels between the model prediction value and the true value.

[0123] In training step 2, the knowledge distillation training method is adopted, using a high-precision wireless propagation model as the teacher model and a transformer network model as the student model. The process of training the recommendation module includes:

[0124] Construct the second training set: randomly crop bounding boxes of different sizes in the building distribution map, and obtain N candidate base station sites through equal-interval sampling; calculate the radio map corresponding to each base station site through the wireless propagation model; based on the radio map, calculate the ranking matrix corresponding to each base station site with the maximization of the available signal coverage as the standard; use the building distribution map and the candidate base station sites obtained based on the bounding box as model input, and use the ranking matrix as training labels;

[0125] The second training set is used to train the recommendation module and the first feedforward network: first, the building distribution map is extracted through the backbone network and the image feature encoder to obtain the image embedding vector of the building distribution map, and the N candidate base station sites are obtained through the prompt encoder to obtain the corresponding regional embedding vectors; the image embedding vector of the building distribution map and the regional embedding vectors of the N base station sites to be deployed are interactively modeled through the image decoder to obtain N image feature codes and base station location codes corresponding to the N candidate base station sites;

[0126] The base station location code is transformed through the first feedforward network and then passes through the recommendation module to obtain the score matrix

[0127] After calculating the ListMLE loss function for the score matrix and the ranking matrix, the parameters of the recommendation module and the first feedforward network are updated with gradients.

[0128] Using ListMLE as the loss function makes the sorting probability distribution output by the transformer network model approximate the true probability distribution L. Its calculation formula is as follows:

[0129]

[0130] Where x represents the building map input to the model, f(x) represents the score output by the model, and y represents the order of the true labels. (j) represents the probability of the jth number in the score matrix; x (k) represents the probability of the kth number in the score matrix. n represents the number of numbers in the score matrix. By maximizing the probability P(y|x; f), the predicted result can be made as close to the true result as possible. During the training process, the transformer network model automatically updates the parameters of the model through the gradient descent algorithm to reduce the difference between the predicted value and the true value.

[0131] All models of the present invention, including pre-trained models, are trained end-to-end. This specific embodiment uses the Adam optimizer to train the model, with the batch size set to 64, the learning rate set to 10-4, and the epoch set to 80.

[0132] The advantages of the present invention are further illustrated below in conjunction with test data.

[0133] This specific embodiment uses the open source dataset RadioMapSeer for training. The dataset contains building maps of 700 cities, which are taken from open street maps of 6 different cities. Each building map is equipped with 80 omnidirectional antenna base stations. And using intelligent ray tracing technology (IRT), 56,000 corresponding high-precision radio maps are obtained.

[0134] The building map and radio map are Min-Max normalized to eliminate the dimension effect and reduce the computational overhead. Then, data enhancement strategies such as image rotation are used to further enhance the robustness of model training. Finally, the building distribution map, base station coordinates, and radio map are scaled to (256*256) as the input and output of the model.

[0135] During the dataset construction process, cells numbered 1-500 were selected as the training set for training the model. Cells numbered 501-600 were selected as the validation set to evaluate the convergence effect of the model during the training process. Cells numbered 601-700 were used as the test set to evaluate the accuracy of different radio estimation models and their generalization capabilities in real scenarios.

[0136] This specific embodiment uses the root mean square error (RMSE) and the normalized mean absolute error (NMSE) to evaluate the performance of the model in the RM estimation task.

[0137] This experiment compares and analyzes the present invention and other advanced RM estimation models. The comparison methods include:

[0138] (1) RadioUnet: This method is the baseline of this paper. It models the underlying physical characteristics of radio propagation through Unet and constructs the mapping relationship between building maps, base stations and RMs.

[0139] (2) TranUnet: This method combines the advantages of self-attention mechanism and Unet to achieve more powerful feature representation.

[0140] (3) RadioNet: This method introduces the Transformer structure into the RM estimation task, significantly improving the accuracy of radio map drawing.

[0141] (4) RadioCycle: This method uses sparse RSRP sampling values ​​to simultaneously restore the building map and RM, and introduces the RSRP measurement truth to calibrate the radio map, which effectively improves the accuracy of model prediction.

[0142] Table 1 Comparison of experimental results of the present invention with other models

[0143]

[0144]

[0145] Table 1 shows the key progress of the proposed method in terms of computational time consumption. The single time consumption represents the running time of estimating the radio map based on a single base station. The total time consumption represents the total running time of the method for recommending the best location among N candidate points (where N=64) in the base station candidate box. The results show that due to the large number of process parameters for optimizing base station site selection, a single predicted radio map takes a long time. However, when running 64 candidate points, the total running time is significantly reduced from more than 1 second to 0.197 seconds, reducing the computational time by 80.3% on the basis of a slight improvement in RM estimation accuracy. This is because the baseline method requires an exhaustive selection of candidate base station locations with a time complexity of O(NM) from N candidate cells in M ​​cells. The present invention reduces the complexity to O(M) by calculating all candidate locations in parallel. Compared with the benchmark method in the prior art, the present invention greatly improves the efficiency of base station site selection in continuous urban space. As Figure 5 As shown, the suffixes train and vaild represent the training set and validation set corresponding to each model, respectively. Compared with the models in the prior art, the present invention has the best effect on the validation set, that is, it has a stronger generalization ability and can better simulate the underlying physical laws of radio wave propagation.

[0146] In the process of selecting the optimal base station location using the trained models and the maximum average RSRP value of the entire image as the standard, methods such as RadioUnet need to select m candidate base station locations with a fixed step size of s=4 when confirming the optimal base station location of n images. Then, the candidate base station locations and the building map are used to cyclically calculate the entire radio map of the current map. However, the prompt encoder of the present invention encodes the entire area, and only one inference process is required to obtain all high-definition radio maps in an area. In order to further demonstrate the performance of the method proposed in the present invention, the candidate points in candidate boxes of different sizes are divided at intervals of 4 meters. The inference time of different models is as follows: Figure 6 As shown in the figure, it is obvious that since the current model adopts a serial radio map estimation mode, the time consumption of the base station recommendation task increases exponentially with the increase of the candidate box. The time consumption of the present invention (radiotransformer) as the number of candidate points in the candidate box increases is significantly less than that of other models.

[0147] An example of practical application is Figure 7 As shown, the base map is a building distribution map, where the red box is the annotation box for the user to select the base station candidate area, and the blue point is the randomly selected base station candidate location. For a fixed number of candidate points in an urban cell, the method proposed in the present invention allows the user to define a base station candidate box of any size and any position. The method proposed in the present invention can quickly recommend a predefined number (64 in the example) of base station locations in the candidate box to maximize the signal coverage. By reducing computational overhead, it can help shorten the running time of base station deployment tasks in urban areas.

[0148] The contents not described in detail in this specification belong to the prior art known to professional and technical personnel in this field.

Claims

1. A base station optimization site selection method based on transformer network, characterized in that: The following steps are involved: Obtain a building distribution map of the expected radiation area of ​​the base station to be deployed; Annotate candidate areas of base stations using bounding boxes in the building distribution map; The backbone network extracts the features of the building distribution map and inputs them into the image feature encoder, and outputs the image embedding vector of the building distribution map to the image decoder; Input the candidate region to the hint encoder, and output the region embedding vector of the candidate region to the image decoder; The image decoder arranges the image embedding vector and the region embedding vector in order, learns the dependency between them, and outputs the radio feature vector and the base station location code. The first feedforward network transforms the base station location encoding dimension and outputs it to the recommendation module; The recommendation module takes the maximization of available signal coverage as the standard and outputs the optimal base station coordinates and their corresponding radio feature vectors based on the base station location coding after dimension transformation; The second feed-forward network generates a weight matrix based on the base station position code; The feature pyramid module generates a radio map corresponding to the optimal base station based on the radio feature vector corresponding to the optimal base station, the weight matrix and the features of the building distribution map; The training process of the transformer network includes: first freezing the recommendation module and the first feedforward network, training the backbone network, the image feature encoder, the prompt encoder, the image decoder, the second feedforward network and the feature pyramid module, so that the transformer network model has the ability to generate a high-definition radio map; then freezing the backbone network, the image feature encoder, the prompt encoder, the image decoder, the second feedforward network and the feature pyramid module, training the recommendation module and the first feedforward network, so that the transformer network model has the ability to perceive the optimal base station; The process of training the backbone network, image feature encoder, hint encoder, image decoder, second feedforward network, and feature pyramid module includes: Constructing a first training set; a single sample of the first training set includes a building distribution map and base station coordinates as model inputs, and a corresponding real radio map as a training label; The backbone network, the image feature encoder, the cue encoder, the image decoder, the second feedforward network and the feature pyramid module are trained using the first training set; the feature pyramid module outputs a reconstructed radio map; After calculating the loss function between the reconstructed radio map and the real radio map, the gradient updates the parameters of the backbone network, image feature encoder, hint encoder, image decoder, second feedforward network and feature pyramid module; The process of training the recommendation module includes: Construct the second training set: randomly select bounding boxes in the building distribution map, and obtain several candidate base station sites by sampling at equal intervals; calculate the radio map corresponding to each base station site; based on the radio map, calculate the ranking matrix corresponding to each base station site with the maximization of the available signal coverage as the standard; use the building distribution map and the candidate base station sites obtained based on the bounding box as model inputs, and use the ranking matrix as training labels; Using the second training set to train the recommendation module and the first feed-forward network; The image decoder outputs base station position codes corresponding to a number of candidate base station sites; The base station location code is transformed into a score matrix through the first feed-forward network and then passed through the recommendation module; After the loss function is calculated for the score matrix and the ranking matrix, the parameters of the recommendation module and the first feedforward network are updated with gradients.

2. A method according to claim 1, characterized in that: The backbone network initially extracts the features of the building distribution map and outputs a set of image feature matrices. It also divides the image features to obtain a set of image embedding vectors, each of which represents a sub-region mapped to the original image space. The image feature decoder is based on the self-attention mechanism and encodes the set of image embedding vectors through the visual transformer to reduce image redundancy.

3. A method according to claim 1, characterized in that: The prompt encoder evenly scatters points within the bounding box with a set step size to obtain several candidate base station positions; then the position coding vectors of several candidate base stations are fused and superimposed with the position coding of the upper left corner and the lower right corner of the bounding box to obtain the regional embedding vector representing the positions of several candidate base stations.

4. A method according to claim 1, characterized in that: The image decoder calculates the importance weights between the feature vector of each image embedding vector and the feature vector of each region embedding vector through the attention mechanism; based on all pairwise importance weights and the nonlinear ReLU layer, through the multi-layer attention mechanism, the feature vectors of the image embedding vector and the region embedding vector are updated to the radio feature vector and the base station location code.

5. A method according to claim 1, characterized in that: The recommendation module obtains the score matrix corresponding to all candidate base stations under the available signal coverage range standard based on the base station position coding after the dimension transformation; and obtains the radio feature vector corresponding to the optimal base station through the optimal selection operation.

6. A method according to claim 1, characterized in that: The feature pyramid module downsamples the features of the building distribution map; restores the radio feature vector corresponding to the optimal base station into image features; and upsamples the image features layer by layer: The downsampled features of the building distribution map at the corresponding level are superimposed and fused with the image features, and then upsampled. The upsampled features are added and fused with the information of the channel dimension supplemented by pixel shuffle as new image features. The above upsampling process is performed repeatedly until the image features of the original building distribution map size are obtained; The image features of the original building distribution map size are multiplied by the weight matrix to obtain the radio map corresponding to the optimal base station.

7. A base station optimization site selection system based on transformer network, characterized in that: It includes a backbone network, an image feature encoder, a prompt encoder, an image decoder, a first feedforward network, a second feedforward network, a recommendation module, and a feature pyramid module; The backbone network is used to extract the features of the building distribution map and input them into the image feature encoder. The image feature encoder is used to output the image embedding vector of the building distribution map to the image decoder; The hint encoder is used for outputting a region embedding vector of the candidate region to the image decoder according to the candidate region; The image decoder is used to arrange the image embedding vector and the region embedding vector in order, learn the dependency between the two, and output the radio feature vector and the base station location code; The first feed-forward network is used to transform the base station location encoding dimension and then output it to the recommendation module; The recommendation module is used to select the optimal base station coordinates and their corresponding radio feature vectors based on the base station location codes after dimension transformation, taking the maximization of the available signal coverage as the standard; The second feed-forward network is used to generate a weight matrix based on the base station position code; The feature pyramid module is used to generate a radio map corresponding to the optimal base station based on the radio feature vector corresponding to the optimal base station, the weight matrix and the features of the building distribution map; The training process of the transformer network includes: first freezing the recommendation module and the first feedforward network, training the backbone network, the image feature encoder, the prompt encoder, the image decoder, the second feedforward network and the feature pyramid module, so that the transformer network model has the ability to generate a high-definition radio map; then freezing the backbone network, the image feature encoder, the prompt encoder, the image decoder, the second feedforward network and the feature pyramid module, training the recommendation module and the first feedforward network, so that the transformer network model has the ability to perceive the optimal base station; The process of training the backbone network, image feature encoder, hint encoder, image decoder, second feedforward network, and feature pyramid module includes: Constructing a first training set; a single sample of the first training set includes a building distribution map and base station coordinates as model inputs, and a corresponding real radio map as a training label; The backbone network, the image feature encoder, the cue encoder, the image decoder, the second feedforward network and the feature pyramid module are trained using the first training set; the feature pyramid module outputs a reconstructed radio map; After calculating the loss function between the reconstructed radio map and the real radio map, the gradient updates the parameters of the backbone network, image feature encoder, hint encoder, image decoder, second feedforward network and feature pyramid module; The process of training the recommendation module includes: Construct the second training set: randomly select bounding boxes in the building distribution map, and obtain several candidate base station sites by sampling at equal intervals; calculate the radio map corresponding to each base station site; based on the radio map, calculate the ranking matrix corresponding to each base station site with the maximization of the available signal coverage as the standard; use the building distribution map and the candidate base station sites obtained based on the bounding box as model inputs, and use the ranking matrix as training labels; Using the second training set to train the recommendation module and the first feed-forward network; The image decoder outputs base station position codes corresponding to a number of candidate base station sites; The base station location code is transformed into a score matrix through the first feed-forward network and then passed through the recommendation module; After the loss function is calculated for the score matrix and the ranking matrix, the parameters of the recommendation module and the first feedforward network are updated with gradients.

Citation Information

Patent Citations

  • Temporary road path planning method and system based on graph neural network

    CN116182875A

  • Method and network apparatus for generating real-time radio coverage map in wireless network

    US20230037893A1