A wireless coverage manifold calibration method based on a graph self-attention mechanism
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NANJING UNIV OF POSTS & TELECOMM
- Filing Date
- 2026-07-03
- Publication Date
- 2026-08-07
AI Technical Summary
[0007]本发明的目的在于提供一种基于图自注意力机制的无线覆盖流形校准方法，用于解决现有深度学习在城市蜂窝网络覆盖流形估计中存在的在无线覆盖测量数据有限、覆盖信息缺失严重的情况下，如何更准确地估计并校准无线覆盖流形等问题；通过引入覆盖流形全局注意力建模以及关键区域采样反馈机制，使模型在覆盖信息缺失条件下同时具备初始流形生成能力、全局相关区域识别能力和采样反馈校准能力，从而提高无线覆盖流形估计的精度与稳定性
本发明能够减少实测采样数量并提高采样有效性，本发明不是在待测区域内随机选择采样点，也不是按照固定间隔进行均匀采样，而是利用覆盖流形全局建模模型中的注意力权重，确定与整体覆盖分布关联度较高的局部区域作为关键采样区域。由此，采样点优先分布在对覆盖流形变化影响较大的区域，使有限数量的实测样本能够提供更多有效校准信息，从而降低不必要的采样开销。
Smart Images
Figure CN122534462A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of intelligent modeling and coverage estimation technology for wireless communication networks. Specifically, it relates to a wireless coverage manifold calibration method based on graph self-attention mechanism, which can be applied to coverage manifold estimation and calibration of cellular communication networks, wireless channel spatial feature estimation, and related intelligent network optimization scenarios. Background Technology
[0002] With the rapid development of cellular mobile communication networks, the Internet of Things (IoT), and smart terminal services, network coverage quality has a significant impact on the planning, optimization, and operation and maintenance of wireless communication systems. Wireless coverage information reflects the spatial distribution of signal strength, signal-to-interference-plus-noise ratio (SIR), reference signal received power, or coverage probability within a target area, and is an important basis for base station deployment, network optimization, interference analysis, and ensuring user experience.
[0003] Traditional methods for acquiring wireless coverage mainly rely on field drive tests, user terminal measurement reports, or simulation calculations based on propagation models. Among these, field drive tests can obtain relatively realistic coverage data, but they require significant manpower, time, and equipment costs, and are difficult to conduct high-density, long-term continuous measurements over large areas. While user terminal measurement reports can reduce measurement costs to some extent, their sampling locations are affected by user distribution, easily leading to problems such as uneven spatial distribution and sparse samples in local areas. Simulation methods based on propagation models can generate coverage estimation results relatively quickly, but their accuracy depends on the accuracy of terrain, buildings, base station parameters, and propagation environment modeling, and are prone to significant errors in complex urban environments.
[0004] In recent years, with the development of deep learning technology, wireless coverage prediction methods based on neural networks have gradually attracted attention. These methods typically take information such as base station location, terrain elevation, building distribution, or partially known coverage values as input, and use models such as convolutional neural networks (CNNs) and U-Net to learn the mapping relationship between environmental features and wireless coverage distribution, thereby estimating the coverage situation in unknown areas. For example, CNN architectures offer significant performance improvements compared to stochastic geometry methods, but the effectiveness still needs further refinement. Works represented by cGAN use adversarial training mechanisms to improve the accuracy of coverage manifold estimation by enhancing the alignment of the generator's mapping of environmental features and coverage manifold features. However, considering the significant differences in features between the environment and coverage manifold across different domains, this supervised training-based cross-domain mapping still cannot meet application requirements in terms of stability and generalization.
[0005] Compared to traditional propagation models, deep learning methods possess stronger nonlinear modeling capabilities, enabling them to learn the spatial distribution patterns of coverage under complex propagation environments from a large number of samples. However, existing deep learning-based wireless coverage prediction methods still have certain shortcomings: On the one hand, models relying solely on convolutional structures focus more on extracting local neighborhood features, limiting their ability to model the global correlations between distant regions in the wireless coverage distribution; on the other hand, in practical applications, the available coverage measurements are often incomplete and unevenly distributed. If the sampling points are not selected appropriately, a small amount of measurement data is insufficient to effectively calibrate the overall coverage distribution. Furthermore, while some methods can generate initial coverage prediction results, they lack a mechanism to actively select key sampling areas based on the global coverage structure and provide feedback to correct the prediction results. This leads to problems such as coverage boundary shifts, large local errors, and instability of the overall manifold structure in prediction results under high missing rates or complex propagation scenarios.
[0006] Therefore, how to fully utilize the global structural information of the wireless coverage manifold under limited sampling conditions, determine the key sampling areas that have a significant impact on the overall coverage distribution, and feed the sampling results back to the coverage prediction process in order to achieve effective calibration of the wireless coverage manifold is a technical problem that urgently needs to be solved in the field of wireless coverage prediction and network optimization. Summary of the Invention
[0007] The purpose of this invention is to provide a wireless coverage manifold calibration method based on a graph self-attention mechanism, which addresses the problems in existing deep learning methods for estimating and calibrating wireless coverage manifolds in urban cellular networks, particularly when wireless coverage measurement data is limited and coverage information is severely lacking. By introducing global attention modeling of the coverage manifold and a key area sampling feedback mechanism, the model simultaneously possesses the ability to generate initial manifolds, identify globally relevant areas, and perform sampling feedback calibration under conditions of missing coverage information, thereby improving the accuracy and stability of wireless coverage manifold estimation.
[0008] To achieve the above objectives, the technical solution of the present invention is implemented as follows: A wireless coverage manifold calibration method based on graph self-attention mechanism includes the following steps: Step 1: Obtain a wireless coverage sample dataset and train a coverage manifold prediction model based on the wireless coverage sample dataset; wherein, the wireless coverage sample dataset includes base station location samples, terrain height samples and corresponding wireless coverage manifold samples, and the coverage manifold prediction model is used to generate a wireless coverage manifold based on the base station location, terrain height and input coverage sampling information; Step 2: Construct a global modeling model for the covering manifold. The model sequentially includes an input processing module, an attention modeling module, and a global representation output module. The input processing module performs local region partitioning, feature encoding, and spatial location encoding on the input wireless covering manifold to obtain multiple local region features. The attention modeling module performs global correlation modeling on each local region feature based on a graph self-attention mechanism, extracts the global structural features of the covering manifold, and outputs the attention weights corresponding to each local region. The global representation output module maps the global structural features into a global feature representation that characterizes the overall structure of the covering manifold. Step 3: Train the global modeling model of the covering manifold using a teacher-student comparative learning method, so that the model learns the global structural features of the wireless covering manifold and the correlation between local regions and the global covering manifold; Step 4: Input the trained global modeling model of the wireless coverage manifold to be calibrated, calculate the correlation between the local region and the global coverage manifold according to the attention weights corresponding to each local region, and select the local regions whose correlation meets the preset conditions as key sampling regions. Step 5: Obtain the measured coverage sampling value within the key sampling area, feed the coverage sampling value back to the coverage manifold prediction model, perform secondary calibration on the wireless coverage manifold to be calibrated, and obtain the final calibrated wireless coverage manifold.
[0009] Furthermore, in step 1, the wireless coverage sample dataset includes publicly available datasets, historical measurement datasets, or simulation-generated datasets; the coverage area size of each wireless coverage sample is [size missing]. Among them, base station location samples A binary matrix Terrain height sample A real matrix Wireless coverage manifold samples Coverage probability matrix All samples are divided into training and test sets that do not overlap, according to a preset ratio.
[0010] Furthermore, in step 1, the coverage manifold prediction model has two capabilities: when a full-mask coverage manifold is input, it generates a complete initial wireless coverage manifold based solely on base station location samples and terrain height samples; when a coverage manifold containing partial measured values is input, it reconstructs and calibrates the missing areas based on known sampling points.
[0011] Furthermore, in step 2, the specific operations of the input processing module include: performing at least two data augmentations on the input wireless coverage manifold sample, with augmentation methods including rotation, flipping, small Gaussian noise perturbation, low-frequency smoothing perturbation, and random local occlusion, to obtain multiple different augmented views; and discretizing each augmented view into a resolution of [resolution missing]. The non-overlapping local regions are flattened into vectors to obtain... Local region vectors, where For the side length of the coverage area, Let be the side length of the local region; the local region vector is mapped to a high-dimensional feature through a learnable projection matrix, and learnable global vectors and position vectors are added to obtain the feature encoding representation of the input covering manifold.
[0012] Furthermore, in step 2, the attention modeling module is composed of... The Transformer coding blocks are stacked, and each Transformer coding block contains a layer normalization unit, a multi-head self-attention unit, a residual connection unit, and a feedforward network unit in sequence. The correlation between features of different local regions is calculated through the multi-head self-attention mechanism, so that each local region feature is updated in combination with global information, and finally the global structural features covering the manifold are output.
[0013] Furthermore, in step 2, the global representation output module is a global representation head connected after the Transformer backbone network. It consists of a multi-layer fully connected network, a batch normalization layer, and a non-linear activation function in sequence, and is used to map the global features output by the Transformer into feature vectors required for contrastive learning.
[0014] Furthermore, in step 3, the specific implementation of the teacher-student comparative learning method is as follows: the teacher model and the student model adopt the same global modeling structure of the covering manifold; two different augmented views are generated for the same wireless covering manifold sample, the unmasked augmented view is input into the teacher model, and the augmented view after random local occlusion is input into the student model; a bidirectional comparative learning loss is constructed based on the global feature vectors output by the teacher model and the student model, so that the features of different augmented views of the same sample are close to each other in the feature space, and the features of different samples are distinguished from each other.
[0015] Furthermore, in step 3, the bidirectional contrastive learning loss function is:
[0016] in, For batch size, For dot product operation, These are the global feature vectors output by the teacher model and the student model, respectively, and belong to different augmented views. Temperature coefficient; The teacher model parameters are not included in gradient backpropagation; they are updated based on the student model parameters using an exponential moving average method.
[0017] in, Update the weights for the parameters of the teacher model. and These are the parameters for the student model and the teacher model, respectively.
[0018] Furthermore, in step 4, the method for determining the key sampling region is as follows: Take the trained student model, input the wireless coverage manifold model to be calibrated, and extract the multi-head attention matrix of the last Transformer coding block; right The final attention matrix is obtained by averaging the attention matrices of the individual attention heads. The dimension of each attention head is , For feature projection dimension; Take the vector after removing the first element from the first row of the matrix. Each element corresponds to the correlation between a local region and the global covering manifold; right Sort in descending order and select the top-ranked items. Several local areas were selected as key sampling areas, among which The number of key areas is preset.
[0019] Furthermore, in step 5, the specific operations of the secondary calibration include: Determine the set of sampling points based on the key sampling area. Construct the sampling position matrix ,in Indicates position Participate in measurement, Indicates position Not involved in the measurement; Obtain the measured coverage values of key sampling areas and construct a sample of the measured coverage manifold. ,in This represents element-wise multiplication in a matrix; Base station location samples Terrain height samples and measured coverage manifold samples Input the coverage manifold prediction model, output the final calibrated wireless coverage manifold. .
[0020] Beneficial effects: This invention reduces the number of measured samples and improves sampling effectiveness. Instead of randomly selecting sampling points within the test area or performing uniform sampling at fixed intervals, it utilizes the attention weights in the global manifold model to identify key sampling areas that are highly correlated with the overall coverage distribution. Therefore, sampling points are preferentially distributed in areas that significantly influence changes in the coverage manifold, allowing a limited number of measured samples to provide more effective calibration information and thus reducing unnecessary sampling overhead.
[0021] This invention can improve the calibration accuracy of covered manifolds: The invention backfills the measured coverage values of key sampling areas into the covered manifold to be calibrated, and then inputs them again into the covered manifold prediction model for secondary inference. This allows the model to use a small amount of real measurement information to correct the initial prediction results, thereby reducing the covered manifold estimation error and improving the prediction accuracy of the calibrated covered manifold. Attached Figure Description
[0022] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a flowchart illustrating the overall steps of the wireless coverage manifold calibration method based on graph self-attention mechanism of the present invention. Figure 2 This is an architecture diagram of the global manifold modeling model of this invention; Figure 3 This is a flowchart illustrating the training process of the global manifold modeling model of this invention. Figure 4 This is a flowchart illustrating the deployment process of the present invention for determining key sampling regions based on an attention mechanism. Figure 5 This is a flowchart illustrating the wireless coverage manifold calibration based on key area sampling feedback according to the present invention. Figure 6 This is an experimental comparison of the mean absolute error between the method of this invention and the random sampling calibration method. Detailed Implementation
[0023] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.
[0024] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.
[0025] The wireless coverage manifold calibration method based on graph self-attention mechanism disclosed in this embodiment is divided into an offline model training stage and an online deployment calibration stage. The offline training stage corresponds to the attached... Figure 3 The online deployment phase corresponds to the attached Figure 4 and attached Figure 5 This embodiment achieves high-precision and high-efficiency wireless coverage manifold estimation and calibration under conditions of missing coverage information and sparse sampling points through a closed-loop mechanism of "initial prediction - global analysis - key sampling - feedback calibration".
[0026] Example 1 See Figure 1 Specifically, it includes the following steps: Step 1: Obtain the wireless coverage sample dataset and train the coverage manifold prediction model. Obtain a wireless coverage sample dataset, which may be a publicly available dataset, a historical measurement dataset, or a simulation-generated dataset. The dataset contains... There are [number] wireless coverage samples, each sample covering an area of [size]. The regular grid, each wireless coverage sample includes: Base station location samples , is a binary matrix. The element 1 indicates that a base station exists in the corresponding grid, and 0 indicates that no base station exists. Terrain height sample , is a real matrix The element value represents the altitude of the corresponding raster cell; Corresponding wireless coverage manifold samples , is the coverage probability matrix, The element value represents the probability of wireless signal coverage for the corresponding grid.
[0027] All the above samples are divided into training and test sets according to a preset ratio (e.g., 7:3), ensuring that the two sets do not overlap. The training set is used for model parameter learning, and the test set is used for model performance validation.
[0028] A manifold prediction model was trained based on the aforementioned dataset. ,in These are the model parameters. The covering manifold prediction model possesses two core capabilities: 1. When the input is a full masked coverage manifold, the complete initial wireless coverage manifold is generated only based on base station location samples and terrain height samples; 2. When the input is a covered manifold containing some measured values, the missing regions are reconstructed and calibrated based on known sampling points.
[0029] Step 2: Construct a global modeling model of the covering manifold The architecture of the global modeling model covering the manifold is shown in the attached figure. Figure 2 As shown, it includes an input processing module, an attention modeling module, and a global representation output module, in that order. The specific construction process is as follows: Step 2.1: Input Processing Module For the received covering manifold samples To simplify the explanation, let... .
[0030] First of all Two data augmentations were performed, each involving rotation, flipping, small Gaussian noise perturbation, low-frequency smoothing perturbation, and random local occlusion to form the input. Two different enhanced views and .
[0031] Step 2.1.1: Local Region Division Input will be displayed under different enhanced views. Discretize to a resolution of By flattening the non-overlapping local regions into vectors, we obtain: (1) in, The total number of local areas. For the corresponding number Flattening vectors for local regions.
[0032] Step 2.1.2: Feature Encoding and Spatial Location Encoding The local region vector is linearly projected through a learnable projection matrix, and then a learnable global vector and a position vector are added to obtain the feature encoding representation of the input covering manifold: (2) in, These are artificially added, learnable global vectors used to aggregate global features; It is a learnable projection matrix; It is a learnable location vector used to preserve the spatial location information of local regions; Indicates the dimension of the feature projection; The feature encoding representation of the input covering manifold.
[0033] Step 2.2: Attention Modeling Module The input features obtained in step 2.1 Input by An attention modeling module composed of stacked Transformer encoding blocks. The calculation process of a layer Transformer coded block is represented as follows: (3) (4) in, ; This indicates a multi-head self-attention module; Representation layer normalization; Indicates a feedforward network; Indicates the first Feature representation of the layer output. After... After processing by the Transformer encoding block, the first row of the output features is extracted as the global feature representation, and then layer normalization is performed: (5) in, Indicates the first The first row of the layer's output features is the global feature representation;
[0034] This represents the normalized global feature representation.
[0035] Step 2.3: Global Representation Output Module A global representation head is connected after the Transformer backbone network. This global representation head consists of multiple fully connected layers, batch normalization layers, and non-linear activation functions, used to process global features. The mapping is performed to the feature vectors required for contrastive learning. The calculation process is as follows: (6) (7) (8) in, Linear layer; For batch normalization layer; It is a non-linear activation function; This represents the feature representation output by the intermediate layer. This refers to the hidden dimensions of the intermediate layer. This is the final output global feature representation. This refers to the output layer dimension.
[0036] Step 3: Train the global modeling model covering the manifold using a teacher-student comparative learning approach. The training process is as follows: Figure 3 As shown, a contrastive learning approach based on a teacher-student structure is used to train the global modeling model covering the manifold. The teacher model and the student model use the same structure, both using Vision Transformer as the backbone network and outputting global feature vectors through a global representation head.
[0037] For the same wireless coverage manifold sample Enhanced view without obscuring and Input the teacher model, where, Enhanced view after random partial occlusion and Input student model, where, During training, a bidirectional contrastive learning loss is constructed based on the global feature vectors output by the teacher model and the student model, so that different augmented views of the same sample remain consistent in the feature space, and the feature representations of different samples remain distinguishable.
[0038] The contrastive learning loss function of the student model is calculated, and the student model parameters are updated through backpropagation. The loss function used is the InfoNCE loss, and its expression is: (9) in, Batch size; This represents the dot product operation; These are different augmented views of global feature vectors output by the teacher model and the student model, respectively, and belonging to the same wireless coverage manifold sample; This is a temperature coefficient used to adjust the smoothness of the characteristic distribution.
[0039] The teacher model parameters are not included in the gradient backpropagation of the loss function. Instead, they are updated using an exponential moving average based on the student model parameters to improve the model's robustness to manifold perturbations and view changes. The update formula is: (10) in, Update the weights for the parameters of the teacher model; For the parameters of the student model; These are the parameters of the teacher model.
[0040] Step 4: Determine the key sampling region based on the attention mechanism Based on the attention weights between global features and local region features, the relevance of each local region to the global coverage manifold is characterized; that is, the relevance between each local region and the global wireless coverage manifold is calculated based on the attention weights, and local regions whose relevance meets the preset conditions are selected as key sampling regions from multiple local regions according to the relevance sorting, threshold filtering or preset quantity filtering.
[0041] During deployment, the wireless coverage manifold to be calibrated is first generated, then input into the trained student model, and key sampling regions are determined based on attention weights for subsequent coverage sampling and calibration; specific steps are attached. Figure 4 As shown: Step 4.1: Generate fully masked manifold samples Take test set base station location samples Terrain height samples and wireless coverage manifold samples Perform a full masking operation on the wireless coverage manifold sample to mask all coverage information: (11) in, This represents element-wise matrix multiplication. Represents a matrix of all zeros; This represents the overlay manifold sample after full masking.
[0042] Step 4.2: Generate the manifold to be calibrated Input the data samples obtained in step 4.1 into the trained covering manifold prediction model to generate the covering manifold to be calibrated: (12) in, The manifold to be calibrated; To cover the manifold prediction model; These are the model parameters.
[0043] Step 4.3: Calculate attention weights and select key sampling regions. The manifold to be calibrated Input the data into the trained student model, extract the multi-head attention matrix of the last Transformer encoding block, and calculate the correlation between each local region and the global covering manifold.
[0044] First, according to formulas (1) to (5), take the first... Input of a Transformer coded block , A multi-head self-attention mechanism is employed for it. Assume there are a total of... First, pay attention to the head. Perform three linear transformations to generate the query vector, key vector, and value vector: (13) in, ; The first Each attention head corresponds to a query vector, a key vector, and a value vector; These are the corresponding trainable matrices; Dimensions for each attention head.
[0045] Calculate the attention matrix for each attention head using the query vector and key vector: and to The final attention matrix is obtained by averaging the attention matrices of the individual attention heads. (14) in, This is the final attention matrix. After removing the diagonal elements, each row of this matrix contains an attention matrix. An attention vector of dimension represents the correlation between the corresponding local region features and the global features.
[0046] Extract the first row of the attention matrix (corresponding to the attention weights in the global CLS vector), and remove the first element to obtain the correlation vector between each local region and the global covering manifold: (15) in, ; Indicates the first The correlation between a local region and the global covering manifold.
[0047] For the correlation vector Sort in descending order to obtain the sorted index sequence: (16) in, This represents the sequence of local region indices arranged from highest to lowest relevance, satisfying the following: (17) Select the top-ranked A local area is designated as the key sampling area. The preset number of key regions (usually a small value is chosen to control sampling costs): (18) in, Represents the set of key sampling regions. Indicates the sorted order of the first... A local area, Generally, the cost is very small, meaning the actual measurement cost is also very small.
[0048] Step 5: Calibrate the wireless coverage manifold based on key area sampling feedback During deployment, on-site measurements are taken in key sampling areas, and the measured values are fed back to the coverage manifold prediction model for secondary calibration to obtain the final wireless coverage manifold. Specific steps are attached. Figure 5 As shown: Step 5.1: Based on the key sampling area obtained in Step 4, determine the set of sampling points: (19) in, This represents the set of locations corresponding to the key sampling region. This represents the set of sampling points selected within the critical sampling region.
[0049] Step 5.2: Construct the sampling location matrix: Construct a binary sampling position matrix Used to mark the location of sampling points: (20) in, Represents the sampling position matrix. Indicates position Participate in measurement, Indicates position Not involved in the measurement.
[0050] Step 5.3: Obtain the measured coverage manifold sample: Field measurements were conducted within the key sampling area to obtain measured coverage values, and a sample of the measured coverage manifold was constructed. (twenty one) in, This represents element-wise multiplication in a matrix. This indicates that the measured coverage manifold sample has a measured value only at the sampling point location, and 0 at other locations.
[0051] Step 5.4: Generate the calibrated wireless coverage manifold Base station location samples Terrain height samples and measured coverage manifold samples The data is input into the coverage manifold prediction model, and secondary inference is performed to generate the final calibrated wireless coverage manifold. (twenty two) in, This is the predicted wireless coverage manifold after final calibration; the predicted wireless coverage manifold after calibration significantly improves performance compared to the coverage manifold to be calibrated.
[0052] The experimental results are attached. Figure 6 As shown, with the increase of the proportion of critical sampling points, the mean absolute error (MAE) of the method of the present invention decreases significantly faster than that of the random sampling calibration method, and maintains a lower error at the same sampling ratio. When the sampling point ratio is 5%, the MAE of the method of the present invention is about 10% lower than that of the random sampling method, verifying the excellent calibration performance of the present invention under limited sampling conditions.
[0053] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A wireless coverage manifold calibration method based on graph self-attention mechanism, characterized in that, Includes the following steps: Step 1: Obtain a wireless coverage sample dataset and train a coverage manifold prediction model based on the wireless coverage sample dataset; wherein, the wireless coverage sample dataset includes base station location samples, terrain height samples and corresponding wireless coverage manifold samples, and the coverage manifold prediction model is used to generate a wireless coverage manifold based on the base station location, terrain height and input coverage sampling information; Step 2: Construct a global modeling model for the covering manifold. The model sequentially includes an input processing module, an attention modeling module, and a global representation output module. The input processing module performs local region partitioning, feature encoding, and spatial location encoding on the input wireless covering manifold to obtain multiple local region features. The attention modeling module performs global correlation modeling on each local region feature based on a graph self-attention mechanism, extracts the global structural features of the covering manifold, and outputs the attention weights corresponding to each local region. The global representation output module maps the global structural features into a global feature representation that characterizes the overall structure of the covering manifold. Step 3: Train the global modeling model of the covering manifold using a teacher-student comparative learning method, so that the model learns the global structural features of the wireless covering manifold and the correlation between local regions and the global covering manifold; Step 4: Input the trained global modeling model of the wireless coverage manifold to be calibrated, calculate the correlation between the local region and the global coverage manifold according to the attention weights corresponding to each local region, and select the local regions whose correlation meets the preset conditions as key sampling regions. Step 5: Obtain the measured coverage sampling value within the key sampling area, feed the coverage sampling value back to the coverage manifold prediction model, perform secondary calibration on the wireless coverage manifold to be calibrated, and obtain the final calibrated wireless coverage manifold.
2. The wireless coverage manifold calibration method based on graph self-attention mechanism according to claim 1, characterized in that, In step 1, the wireless coverage sample dataset includes publicly available datasets, historical measurement datasets, or simulation-generated datasets; the coverage area size of each wireless coverage sample is [size missing]. Among them, base station location samples A binary matrix Terrain height sample A real matrix Wireless coverage manifold samples Coverage probability matrix All samples are divided into training and test sets that do not overlap, according to a preset ratio.
3. The wireless coverage manifold calibration method based on graph self-attention mechanism according to claim 1, characterized in that, In step 1, the coverage manifold prediction model has two capabilities: when a full-mask coverage manifold is input, it generates a complete initial wireless coverage manifold based only on base station location samples and terrain height samples; when a coverage manifold containing partial measured values is input, it reconstructs and calibrates the missing areas based on known sampling points.
4. The wireless coverage manifold calibration method based on graph self-attention mechanism according to claim 1, characterized in that, In step 2, the specific operations of the input processing module include: performing at least two data augmentations on the input wireless coverage manifold sample, with augmentation methods including rotation, flipping, small Gaussian noise perturbation, low-frequency smoothing perturbation, and random local occlusion, to obtain multiple different augmented views; and discretizing each augmented view into a resolution of [resolution missing]. The non-overlapping local regions are flattened into vectors to obtain... Local region vectors, where For the side length of the coverage area, Let be the side length of the local region; the local region vector is mapped to a high-dimensional feature through a learnable projection matrix, and learnable global vectors and position vectors are added to obtain the feature encoding representation of the input covering manifold.
5. The wireless coverage manifold calibration method based on graph self-attention mechanism according to claim 1, characterized in that, In step 2, the attention modeling module is composed of The Transformer coding blocks are stacked, and each Transformer coding block contains a layer normalization unit, a multi-head self-attention unit, a residual connection unit, and a feedforward network unit in sequence. The correlation between features of different local regions is calculated through the multi-head self-attention mechanism, so that each local region feature is updated in combination with global information, and finally the global structural features covering the manifold are output.
6. The wireless coverage manifold calibration method based on graph self-attention mechanism according to claim 1, characterized in that, In step 2, the global representation output module is a global representation head connected after the Transformer backbone network. It consists of a multi-layer fully connected network, a batch normalization layer, and a non-linear activation function in sequence. It is used to map the global features output by the Transformer into feature vectors required for contrastive learning.
7. The wireless coverage manifold calibration method based on graph self-attention mechanism according to claim 1, characterized in that, In step 3, the specific implementation of the teacher-student contrastive learning method is as follows: the teacher model and the student model adopt the same global modeling structure of the covering manifold; two different augmented views are generated for the same wireless covering manifold sample, the unmasked augmented view is input into the teacher model, and the augmented view after random local occlusion is input into the student model; a bidirectional contrastive learning loss is constructed based on the global feature vectors output by the teacher model and the student model, so that the features of different augmented views of the same sample are close to each other in the feature space, and the features of different samples are distinguished from each other.
8. The wireless coverage manifold calibration method based on graph self-attention mechanism according to claim 7, characterized in that, In step 3, the bidirectional contrastive learning loss function is: in, For batch size, For dot product operation, These are the global feature vectors output by the teacher model and the student model, respectively, and belong to different augmented views. Temperature coefficient; The teacher model parameters are not included in gradient backpropagation; they are updated based on the student model parameters using an exponential moving average method. in, Update the weights for the parameters of the teacher model. and These are the parameters for the student model and the teacher model, respectively.
9. The wireless coverage manifold calibration method based on graph self-attention mechanism according to claim 1, characterized in that, In step 4, the method for determining the key sampling region is as follows: Take the trained student model, input the wireless coverage manifold model to be calibrated, and extract the multi-head attention matrix of the last Transformer coding block; right The final attention matrix is obtained by averaging the attention matrices of the individual attention heads. The dimension of each attention head is , For feature projection dimension; Take the vector after removing the first element from the first row of the matrix. Each element corresponds to the correlation between a local region and the global covering manifold; right Sort in descending order and select the top-ranked items. Several local areas were selected as key sampling areas, among which The number of key areas is preset.
10. The wireless coverage manifold calibration method based on graph self-attention mechanism according to claim 1, characterized in that, Step 5, the specific operations of the secondary calibration include: Determine the set of sampling points based on the key sampling area. Construct the sampling position matrix ,in Indicates position Participate in measurement, Indicates position Not involved in the measurement; Obtain the measured coverage values of key sampling areas and construct a sample of the measured coverage manifold. ,in This represents element-wise multiplication in a matrix; Base station location samples Terrain height samples and measured coverage manifold samples Input the coverage manifold prediction model, output the final calibrated wireless coverage manifold. .