Contrastive self-supervised wireless coverage manifold prediction and calibration integrated method

Abstract

The application discloses a wireless coverage manifold prediction and calibration integrated method based on contrast self-supervision, and belongs to the field of intelligent modeling of wireless communication networks. In view of the defects of the existing split type scheme, such as multi-model series error transmission, long inference link, low utilization rate of sampling information and the like, the application constructs an integrated neural network framework containing a multi-source input feature fusion module, an attention encoding module, a global representation output module and a coverage manifold prediction reconstruction module, matches a progressive mask rate and an attention mask strategy, adopts a contrast learning loss and a self-supervised reconstruction loss for joint training, and can simultaneously output an initial coverage prediction result and a key sampling area through one forward inference during deployment, and the calibration is completed through a second inference based on actual measurement backfilling. The application eliminates the multi-model error accumulation from the root, and greatly improves the inference efficiency and sampling accuracy.

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 an integrated method for wireless coverage manifold prediction and calibration based on contrastive self-supervision, 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 and U-Net to learn the mapping relationship between environmental features and wireless coverage distribution, thereby estimating the coverage situation in unknown areas. Compared with traditional propagation models, deep learning methods have stronger nonlinear modeling capabilities and can learn the spatial distribution patterns of coverage under complex propagation environments from a large number of samples. However, considering the significant differences in characteristics between different domains of the environment and coverage manifold, supervised training-based cross-domain mapping still cannot meet application requirements in terms of stability and generalization.

[0005] Furthermore, some existing wireless coverage calibration schemes employ a non-integrated (split-type) architecture that separates the prediction model from the global analysis model. This splits initial coverage prediction, key sampling area selection, and measured feedback calibration into independent processing steps, executed by different network models. Such schemes require multiple models to be called sequentially in a fixed order, resulting in long inference links, high computational time, and high system deployment complexity. Simultaneously, prediction errors from preceding models propagate to subsequent models, affecting the selection of subsequent sampling areas and the final calibration accuracy, leading to error accumulation. Due to the lack of end-to-end collaborative optimization mechanisms between different models, deep feature information cannot be shared, resulting in inconsistencies between the coverage prediction results and the feature basis of the sampling decisions. Limited measured sampling data cannot maximize calibration effectiveness, and the error reduction tends to saturate as the number of sampling points increases, leading to low utilization of sampling information. In scenarios with high missing rates and complex terrain occlusion, the stability and accuracy of coverage manifold estimation are insufficient to meet engineering requirements.

[0006] Therefore, how to overcome the limitations of the existing split architecture, construct an end-to-end model that integrates prediction and sampling, eliminate the propagation of errors between multiple models, improve inference efficiency and sampling synergy, enhance the utilization of sampling information, and further improve the accuracy and stability of coverage manifold calibration in complex environments under limited sampling conditions are the technical problems that urgently need 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 an integrated method for wireless coverage manifold prediction and calibration based on contrastive self-supervision, addressing the problems of insufficient global correlation modeling capability, difficulty in determining key sampling regions under finite sampling conditions, and the lack of feedback calibration mechanisms in existing wireless coverage prediction methods, where neural network models typically only output one-time coverage prediction results. This invention constructs an integrated neural network framework for coverage manifold prediction and key sampling region selection, and utilizes a contrastive self-supervision method to simultaneously obtain the wireless coverage manifold prediction result and key sampling region location information during a single forward inference process. The measured coverage probability value of the key sampling region is then fed back to the coverage prediction process, achieving secondary correction and effective calibration of the initial coverage manifold prediction result, 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 contrastive self-supervised method for integrated prediction and calibration of wireless coverage manifolds includes the following steps: Step 1: Construct a wireless coverage sample dataset. Divide the target city area into multiple geographical regions. For each geographical region, obtain terrain height samples and set multiple base station location distributions to generate corresponding wireless coverage manifold samples. Take each set of base station location samples, terrain height samples and corresponding coverage manifold samples as a data sample and divide them into training sets and test sets that do not overlap. Step 2: Construct an integrated neural network framework for coverage manifold prediction and calibration. The neural network sequentially includes a multi-source input feature fusion module, an attention encoding module, a global representation output module, and a coverage manifold prediction and reconstruction module. The multi-source input feature fusion module receives base station location samples, terrain height samples, and wireless coverage manifold samples and fuses them into a unified fused feature sequence. The attention encoding module performs unified attention encoding on the fused feature sequence, outputting an encoded feature sequence containing global structural features and local region features, along with an attention weight matrix. The global representation output module outputs a global feature representation characterizing the overall structure of the coverage manifold based on the global structural features. The coverage manifold prediction and reconstruction module predicts and reconstructs the wireless coverage manifold based on local region features. Step 3: Train the integrated neural network framework using a teacher-student comparative self-supervised method; input data samples into the network, and simultaneously obtain global feature representation and coverage manifold prediction and reconstruction results; construct the total loss function based on the weighted sum of the comparative learning loss function and the self-supervised reconstruction loss function; update the network parameters through backpropagation, so that the network can simultaneously learn the global structural features and local region reconstruction capabilities of the wireless coverage manifold. Step 4: During deployment, the base station location sample, terrain height sample, and full mask coverage manifold sample are input into the trained integrated neural network framework. During one forward inference process, the initial prediction result of the wireless coverage manifold is output simultaneously, and the attention weights corresponding to each local region are extracted. Based on the attention weights, local regions that meet the preset conditions of global relevance are selected 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 integrated neural network framework, perform secondary calibration on the wireless coverage manifold to be calibrated, and obtain the calibrated wireless coverage manifold.

[0009] Furthermore, in step 1, the wireless coverage sample dataset includes There are [number] wireless coverage samples, each sample covering an area of ​​[size]. The base station location samples are real number matrices in the form of Gaussian heatmaps, the terrain height samples are normalized real number matrices, and the wireless coverage manifold samples are coverage probability matrices. All samples are divided into training and test sets according to a predetermined ratio.

[0010] Furthermore, in step 2, the specific operations of the multi-source input feature fusion module include: discretizing the base station location samples, terrain height samples, and wireless coverage manifold samples into samples with a resolution of [resolution value missing]. The non-overlapping local regions are flattened into vectors to obtain Three local region vectors are generated; the local region vectors of the three types of data are mapped to high-dimensional feature codes respectively through a learnable projection matrix; the three types of feature codes are concatenated in dimension and fused into a unified feature sequence through a feedforward network; a learnable global vector and a spatial location vector are added to obtain a fused feature code representation of multi-source input.

[0011] Furthermore, in step 2, the attention encoding 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. In the final output encoded feature sequence, the first row corresponds to the global structure features, and the remaining rows correspond to the features of each local region.

[0012] Furthermore, in step 2: the global representation output module is a global representation head connected after the attention encoding module, consisting of a multi-layer fully connected network, a batch normalization layer, and a non-linear activation function, used to map global structural features into feature vectors required for contrastive learning; the covering manifold prediction and reconstruction module is a prediction and reconstruction head connected after the attention encoding module, consisting of a feedforward network, a rearrangement layer, and a transposed convolutional layer, used to project local region features back to the original spatial resolution to generate a single-channel covering manifold reconstruction result.

[0013] Furthermore, in step 3, the specific implementation of the teacher-student comparative self-supervision method is as follows: The teacher model and the student model adopt the same integrated neural network framework structure; the base station location, terrain height and coverage manifold of the same data sample are simultaneously augmented twice. The augmentation methods include random rotation, random flipping, random cropping and random noise addition, resulting in two different augmented views as input to the teacher model.

[0014] Furthermore, in step 3, the training process employs an attention masking and progressive masking rate strategy: Extract the multi-head attention matrix of the teacher model, calculate the correlation between each local region and the global coverage manifold, and sort them in descending order; A progressive mask rate is adopted: the mask rate gradually increases from 0 to 30% in the early stage of training, from 30% to 70% in the middle stage of training, and gradually increases to 100% full mask in the later stage of training. Based on the current masking rate, select the local area with the lowest relevance as the masking area, construct a masking matrix to mask the coverage manifold of the augmented view, and obtain the masked augmented view as the input of the student model. In order to prevent the student model from making predictions based solely on environmental information during the reconstruction process, base station location samples and terrain height samples are randomly discarded with a fixed probability when inputting.

[0015] Furthermore, in step 3, the total loss function is constructed as follows: 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. A self-supervised reconstruction loss is constructed based on the reconstruction results of the covered manifold output by the student model and the corresponding original samples, which respectively constrains the reconstruction errors of the masked region and the unmasked region. The total loss function is a weighted sum of the contrastive learning loss and the self-supervised reconstruction loss:

[0016] in, To compare learning loss, To compensate for losses in self-monitoring and reconstruction, These are the weighting coefficients; 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, To update the weights, 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 base station location sample, terrain height sample and full mask coverage manifold sample into the model, and output the initial coverage manifold prediction result after one forward inference; Extract the multi-head attention matrix of the last Transformer encoding block, and then... The final attention matrix is ​​obtained by averaging the attention matrices of the individual attention heads. Take the vector after removing the first element from the first row of the matrix as the correlation vector between each local region and the global covering manifold; Select items based on relevance ranking, threshold filtering, or preset quantity filtering, ensuring that the relevance meets the preset criteria. A local area is designated as the key sampling area.

[0019] Furthermore, in step 5, the specific operations for the secondary calibration include: The set of sampling points is determined based on the key sampling area, and a binary sampling position matrix is ​​constructed, where an element of 1 indicates that the corresponding position is a sampling point, and an element of 0 indicates that it is not a sampling point. Based on the real coverage manifold of the test set, the sampling position matrix is ​​multiplied element by element with the coverage manifold of the corresponding region to extract the measured coverage value of the key sampling region, and a sparse measured coverage manifold sample is obtained where only the sampling positions retain valid values ​​and the remaining positions are zero. Base station location samples, terrain height samples, and measured coverage manifold samples are input into the same integrated neural network framework, and the final calibrated wireless coverage manifold is output after secondary inference.

[0020] Beneficial effects: (1) Integration of manifold prediction and key sampling region selection: This invention constructs a neural network framework that integrates manifold prediction and calibration, and simultaneously completes multi-source feature fusion, global correlation modeling, manifold prediction reconstruction, and key sampling region determination in the same network. Unlike existing neural network methods that only output the coverage prediction results, this invention can simultaneously output the initial prediction results of the wireless manifold and the location information of the key sampling region in one forward inference process, so that the prediction results and sampling decisions come from the same feature representation and attention relationship, avoiding the feature inconsistency and error propagation problems caused by multiple models being connected in series, and improving the synergy between the prediction, sampling and calibration processes.

[0021] (2) Reducing the number of measured samples and improving sampling effectiveness: This invention does not randomly select sampling points within the test area, nor does it perform uniform sampling at fixed intervals. Instead, it utilizes the attention weights in the neural network to determine local areas with a high correlation to the overall coverage distribution as key sampling areas. As a result, sampling points are preferentially distributed in areas that have a greater impact on the change in the coverage manifold, enabling a limited number of measured samples to provide more effective calibration information, thereby reducing unnecessary sampling overhead.

[0022] (3) Improve the calibration accuracy of the covering manifold: The present invention backfills the measured coverage values ​​of the key sampling area into the fully covered manifold, and performs secondary inference through the neural network again, so that the model can use a small amount of real measurement information to correct the initial prediction results, thereby reducing the covering manifold estimation error and improving the prediction accuracy of the calibrated covering manifold. Attached Figure Description

[0023] 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 method of the present invention; Figure 2 This is a diagram of the integrated neural network architecture for manifold prediction and calibration covered by the present invention; Figure 3 This is a flowchart illustrating the training process of the integrated neural network of the present invention. Figure 4 This is a flowchart illustrating the joint determination of initial prediction and critical sampling during the deployment phase of this invention. Figure 5 This is a flowchart of the feedback calibration process based on key area sampling in this invention; Figure 6 This is an experimental diagram comparing the errors of the method of the present invention with those of the comparative method. Detailed Implementation

[0024] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0025] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0026] Example 1 The integrated method for wireless coverage manifold prediction and calibration based on contrastive self-supervision disclosed in this embodiment is divided into an offline training phase and an online deployment and calibration phase. The core is to achieve a closed-loop processing of "initial prediction - key sampling - feedback calibration" through a single network end-to-end integrated architecture.

[0027] like Figure 1 As shown, the entire method consists of five core steps, with steps 1 to 3 being the offline model training process, and steps 4 to 5 being the online deployment and calibration process. The specific training flow is as follows: Figure 3 As shown, the process for jointly determining the initial prediction and key sampling during the deployment phase is as follows: Figure 4 As shown, the key area sampling feedback calibration process is as follows: Figure 5 As shown.

[0028] Step 1: Construct a wireless coverage sample dataset The target city area is divided into multiple geographical regions. For each geographical region, a corresponding terrain height sample is obtained, and multiple base station location distributions are set under the same terrain height sample to form multiple base station location samples. For each set of base station location samples and terrain height samples, the corresponding wireless coverage manifold sample is obtained through simulation calculation or field collection. Each set of base station location samples, terrain height samples, and corresponding wireless coverage manifold samples are used as a data sample to form a training set and a test set that can be used for model learning.

[0029] Specifically, the wireless coverage sample dataset contains There are [number] wireless coverage samples, each sample covering an area of ​​[size]. The grid is a regular grid. Each wireless coverage sample includes three types of data: Base station location samples , is a real matrix in Gaussian heatmap form. A heat distribution is generated centered on the location of the base station to characterize the spatial deployment structure of the base station and its surrounding influence range. Terrain height sample , is a normalized real matrix. The element value represents the normalized terrain height of the corresponding raster; Corresponding wireless coverage manifold samples , is the coverage probability matrix, The element value represents the probability that the signal-to-interference-to-noise ratio (SINR) of the wireless signal of the corresponding grid is greater than a preset threshold.

[0030] All the above samples are divided into training and test sets according to a predetermined ratio, ensuring that there is no overlap between the two. The training set is used for model parameter learning, and the test set is used for model performance validation.

[0031] Step 2: Construct an integrated neural network framework for manifold prediction and calibration. Build an integrated neural network framework, with an architecture as follows: Figure 2 As shown, it includes, in sequence, a multi-source input feature fusion module, an attention encoding module, a global representation output module, and a manifold prediction and reconstruction module. The specific construction process is as follows: Step 2.1: Multi-source input feature fusion module For the received base station location samples Terrain height samples and covering manifold samples Perform feature fusion processing.

[0032] Step 2.1.1: Local Region Partitioning. Discretize the three types of input data samples into regions with a resolution of [resolution value missing]. The non-overlapping local regions are flattened into vectors to obtain: (1) (2) (3) in, This represents the total number of units in the local area. These are the base station location samples, terrain height samples, and coverage manifold samples, respectively. Flattening vectors for local regions.

[0033] Step 2.1.2: Multi-source feature encoding maps the three types of local region vectors into high-dimensional features using a learnable projection matrix: (4) (5) (6) in, It is a learnable projection matrix. Indicates the dimension of the feature projection.

[0034] Step 2.1.3: Multi-source feature fusion and spatial consistency coding The three types of encoded features are concatenated dimensionally, fused into a unified feature sequence via a feedforward network, and then a learnable global vector and a position vector are added. (7) (8) in, Indicates a serial connection operation; Indicates a feedforward network; These are artificially added, learnable global vectors used to aggregate global features; It is a learnable location vector that preserves the spatial location information of the local region; It is a fusion feature encoding representation of multi-source inputs.

[0035] Step 2.2: Attention Encoding Module Fusion features Input by An attention coding module composed of stacked Transformer coding blocks. The calculation process for the Transformer coded block is as follows: (9) (10) 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.

[0036] After L layers of encoding, global feature representation and local region feature representation are extracted respectively: (11) (12) in, This represents the first row of the output features of the Lth layer, i.e., the global structural features; For the first The remaining rows of the layer output features are local region features; , These are the corresponding features after layer normalization.

[0037] Step 2.3: The global representation output module is connected to the global representation head after the attention encoding module. It consists of multiple fully connected networks, batch normalization layers, and non-linear activation functions, which map the global features into contrastive learning feature vectors. (13) (14) (15) in, Linear layer; For batch normalization layer; and 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. For the output layer dimension.

[0038] Step 2.4: Coverage manifold prediction and reconstruction module Following the attention encoding module, a prediction reconstruction head is connected. This head consists of a feedforward network and transposed convolutional layers, which project local region features back to the original spatial resolution, thereby generating a single-channel overlay manifold reconstruction result. (16) in, Indicates a feedforward network; This is a transpose rearrangement function used to transpose... Local region features are restored to their two-dimensional spatial location; This is a transposed convolution with both kernel size and stride. ; This is the result of single-channel coverage manifold reconstruction.

[0039] Step 3: Teacher-Student Comparative Self-Supervised Joint Training Training process as follows Figure 3 As shown, an end-to-end joint training of the integrated neural network is performed using a teacher-student structure, with the teacher model and student model having completely identical structures. Both the student and teacher models use VisionTransformer as the backbone network, and output global feature vectors and covered manifold prediction reconstruction results through a global representation head and a covered manifold prediction reconstruction head, respectively.

[0040] Step 3.1: Spatial Consistency Data Enhancement Before training, data augmentation was performed twice simultaneously on base station samples, terrain height samples, and coverage manifold samples. Each data augmentation method included random rotation, random flipping, random cropping, and random noise addition to obtain the same input data samples. Two enhanced views and And used as input for the teacher model.

[0041] Step 3.2: Attention Mask and Progressive Mask Rate During the training process, the teacher model is taken as the first... Input of a Transformer coded block A multi-head attention mechanism is employed. Assume there are a total of... First, pay attention to the head. Perform a three-dimensional linear transformation: (17) in, , The number of attention heads; These are the corresponding query vector, key vector, and value vector; These are the trainable matrices corresponding to Query, Key, and Value, respectively. Dimensions for each attention head.

[0042] Calculate the attention matrix for each attention head using the query vector and key vector: (18) in, For the first Attention weight matrix for each attention head. The function represents the normalization of each row of the matrix. This is a scaling factor used to prevent the gradient from vanishing due to excessively large dot product values.

[0043] The average of multiple heads is taken as the final attention matrix: (19) in, This is the final attention matrix. Excluding the diagonal elements, each row of this matrix contains an attention matrix. The attention vector is dimensional, representing the correlation between the current local region features and the global features. Let's take its first row: (20) in, , Indicates the first The correlation between a local region and the global context.

[0044] according to Sort each local region in descending order of relevance to obtain the sorted index sequence: (twenty one) in, This represents the sequence of local region indices arranged from highest to lowest relevance, satisfying the following: (twenty two) in, .

[0045] Record the mask rate The retention rate , Dynamic changes during training, when A value close to 1 indicates a high mask rate. This indicates a full mask. In the early stages of training, the mask rate gradually increases from 0% to 30%. In the middle stages of training, the mask rate gradually increases from 30% to 70%. In the later stages of training, the mask rate gradually increases to 100%, which is a full mask.

[0046] Select the top-ranked A portion of the local area is reserved, and the rest is used as the mask area: (twenty three) (twenty four) in, Represents the set of reserved regions. Represents the set of mask regions. Indicates the sorted order of the first... A local area.

[0047] Construct a mask matrix based on the set of mask regions. , A value of 0 indicates no masking is performed in the region. Mesh preserved, otherwise A value of 1 indicates that the position is masked. The masking operation is defined as follows: (25) in, This represents element-wise multiplication in a matrix. This represents the masked overlay manifold sample. This represents a matrix consisting entirely of 1s. Take... , As input to the student model.

[0048] During training, a bidirectional contrastive learning loss is constructed based on the global feature vectors output by the student and teacher models. This ensures that different augmented views of the same sample remain consistent in the feature space and that the feature representations of different samples are distinguishable. A self-supervised reconstruction loss is constructed based on the coverage manifold prediction and reconstruction results from the student model and the corresponding augmented views. This allows the neural network framework to learn the mapping relationship between base station location samples, terrain height samples, and mask coverage manifold samples to the complete wireless coverage manifold. The teacher model parameters are updated using an exponential moving average based on the student model parameters, thereby improving the model's robustness to coverage manifold perturbations and view changes.

[0049] Step 3.3: Joint Loss Calculation and Parameter Update Calculate the loss function for the student model and update its parameters through backpropagation. The loss function is as follows: (26) (27) (28) in, For batch size, This represents the dot product operation; These are the global features output by the teacher model and the student model, respectively, and belong to different augmented views. Temperature coefficient; Indicates the first The number of samples with a mask of 1. Represents a matrix consisting entirely of 1s; is the weighting coefficient used to balance the contrastive learning loss and the self-supervised reconstruction loss.

[0050] 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. (29) 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.

[0051] Step 4: Joint Determination of Initial Coverage Prediction and Key Sampling Area Deployment process as follows Figure 4As shown, during deployment, the neural network framework simultaneously outputs wireless coverage manifold prediction results and key sampling area location information during a single forward inference process. Base station location samples, terrain height samples, and full-mask coverage manifold samples are input into the trained neural network framework. Attention weights corresponding to each local region are extracted, and local regions that meet preset conditions for global relevance are selected based on these attention weights as key sampling areas for subsequent coverage sampling and calibration.

[0052] During deployment, the trained student model and test set are used. Base station location samples, terrain height samples, and full-coverage manifold samples are input into the model to obtain the attention weights corresponding to each local region. Based on these attention weights, the correlation between each local region and the global wireless coverage manifold is calculated. Local regions whose correlation meets preset conditions are selected as key sampling regions from multiple local regions according to correlation ranking, threshold filtering, or preset quantity filtering. The specific steps are as follows: Step 4.1: Constructing the Full Mask Input Take test set base station location samples Terrain height sample and wireless coverage manifold samples For wireless coverage manifold samples Perform a full mask operation: (30) in, This represents element-wise multiplication in a matrix. This represents the sample of the covering manifold after full masking. Represents a matrix consisting entirely of zeros.

[0053] Step 4.2: One forward double output, input the data samples obtained in Step 4.1 into the trained student model: (31) in, For initial coverage manifold prediction, For the student model, This represents the parameters of the student model.

[0054] Step 4.3: Selection of Key Sampling Regions The correlation between each local region and the global coverage manifold is calculated based on the attention weights, and local regions with higher correlation are selected as key sampling regions.

[0055] During the deployment phase, the student model is taken. Based on the output features of each Transformer coded block, the top-ranked features are selected according to formulas (17) to (22). A local area is designated as the key sampling area: (32) in, Represents the set of key sampling regions. This represents the m-th local region after sorting, and S represents the preset number of key sampling regions. Generally, the cost is very small, meaning the actual measurement cost is also very small.

[0056] Step 5: Calibrate the wireless coverage manifold based on key area sampling feedback Calibration process as follows Figure 5 As shown, during deployment, coverage sampling values ​​are acquired within the key sampling area and fed back to the neural network framework, enabling the neural network framework to perform secondary calibration on the wireless coverage manifold to be calibrated, thereby obtaining the calibrated wireless coverage manifold.

[0057] During deployment, the key sampling area obtained in step 4 is sampled, and the resulting measurements are input again into the coverage manifold prediction model. The model can then directly generate calibrated coverage manifold information for the corresponding area. The specific steps are as follows: Step 5.1: Determine the set of sampling points Based on the key sampling area obtained in step 4, determine the set of sampling points: (33) 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.

[0058] Step 5.2: Set the sampling position matrix Construct a binary sampling position matrix: (34) in, Represents the sampling position matrix. Indicates position Participate in measurement, Indicates position Not involved in the measurement.

[0059] Step 5.3: Obtain measured values ​​of key sampling areas Take the measured values ​​of key sampling areas and construct a measured coverage manifold sample: (35) in, This represents element-wise multiplication in a matrix. This represents the measured coverage manifold sample.

[0060] The obtained measurements are then input back into the student model, which can directly generate calibrated coverage manifold information for the corresponding region. (36) in, This is the final calibrated manifold prediction. The calibrated manifold prediction shows a significant performance improvement compared to the initial manifold prediction.

[0061] Experimental results are as follows Figure 6 As shown, the mean absolute error of each method generally decreases with the increase of the proportion of key points, indicating that the measured information of the key sampling region can effectively improve the calibration accuracy of the covered manifold. Compared with the random calibration method, the attention calibration method has a lower error at the same proportion of key points, indicating that the key regions selected based on attention weights have a stronger correlation with the global covered manifold and can provide higher-value calibration information.

[0062] Further comparison of the integrated and non-integrated methods reveals that, under attention calibration conditions, the integrated method exhibits a more stable error reduction and the lowest final error. As the proportion of keypoints increases, the error reduction of the non-integrated method gradually approaches saturation, while the integrated method continues to reduce prediction error. This indicates that the integrated framework can more fully integrate the initial prediction, key region selection, and sampling feedback calibration processes, reducing error propagation between stages and improving the consistency between manifold prediction and calibration. These results validate that the attention calibration strategy can improve the effectiveness of key sampling regions and also demonstrate that integrated prediction and calibration modeling enhances the model's ability to utilize feedback sampling information. The combination of these two approaches results in a lower mean absolute error under the same sampling overhead, demonstrating superior calibration accuracy and engineering application value.

[0063] 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 method for integrated prediction and calibration of wireless coverage manifold based on contrastive self-supervision, characterized in that, Includes the following steps: Step 1: Construct a wireless coverage sample dataset. Divide the target city area into multiple geographical regions. For each geographical region, obtain terrain height samples and set multiple base station location distributions to generate corresponding wireless coverage manifold samples. Take each set of base station location samples, terrain height samples and corresponding coverage manifold samples as a data sample and divide them into training sets and test sets that do not overlap. Step 2: Construct an integrated neural network framework for coverage manifold prediction and calibration. The neural network sequentially includes a multi-source input feature fusion module, an attention encoding module, a global representation output module, and a coverage manifold prediction and reconstruction module. The multi-source input feature fusion module receives base station location samples, terrain height samples, and wireless coverage manifold samples and fuses them into a unified fused feature sequence. The attention encoding module performs unified attention encoding on the fused feature sequence, outputting an encoded feature sequence containing global structural features and local region features, along with an attention weight matrix. The global representation output module outputs a global feature representation characterizing the overall structure of the coverage manifold based on the global structural features. The coverage manifold prediction and reconstruction module predicts and reconstructs the wireless coverage manifold based on local region features. Step 3: Train the integrated neural network framework using a teacher-student comparative self-supervised method; input data samples into the network, and simultaneously obtain global feature representation and coverage manifold prediction and reconstruction results; construct the total loss function based on the weighted sum of the comparative learning loss function and the self-supervised reconstruction loss function; update the network parameters through backpropagation, so that the network can simultaneously learn the global structural features and local region reconstruction capabilities of the wireless coverage manifold. Step 4: During deployment, the base station location sample, terrain height sample, and full mask coverage manifold sample are input into the trained integrated neural network framework. During one forward inference process, the initial prediction result of the wireless coverage manifold is output simultaneously, and the attention weights corresponding to each local region are extracted. Based on the attention weights, local regions that meet the preset conditions of global relevance are selected 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 integrated neural network framework, perform secondary calibration on the wireless coverage manifold to be calibrated, and obtain the calibrated wireless coverage manifold.

2. The integrated method for wireless coverage manifold prediction and calibration based on contrastive self-supervision as described in claim 1, characterized in that, In step 1, the wireless coverage sample dataset contains There are [number] wireless coverage samples, each sample covering an area of ​​[size]. The base station location samples are real number matrices in the form of Gaussian heatmaps, the terrain height samples are normalized real number matrices, and the wireless coverage manifold samples are coverage probability matrices. All samples are divided into training and test sets according to a predetermined ratio.

3. The integrated method for wireless coverage manifold prediction and calibration based on contrastive self-supervision as described in claim 1, characterized in that, In step 2, the specific operations of the multi-source input feature fusion module include: discretizing the base station location samples, terrain height samples, and wireless coverage manifold samples into samples with a resolution of [resolution value missing]. The non-overlapping local regions are flattened into vectors to obtain Three local region vectors are generated; the local region vectors of the three types of data are mapped to high-dimensional feature codes respectively through a learnable projection matrix; the three types of feature codes are concatenated in dimension and fused into a unified feature sequence through a feedforward network; a learnable global vector and a spatial location vector are added to obtain a fused feature code representation of multi-source input.

4. The integrated method for wireless coverage manifold prediction and calibration based on contrastive self-supervision as described in claim 1, characterized in that, In step 2, the attention encoding 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. In the final output encoded feature sequence, the first row corresponds to the global structure features, and the remaining rows correspond to the features of each local region.

5. The integrated method for wireless coverage manifold prediction and calibration based on contrastive self-supervision as described in claim 1, characterized in that, In step 2: the global representation output module is a global representation head connected after the attention encoding module, which consists of a multi-layer fully connected network, a batch normalization layer, and a non-linear activation function in sequence, used to map global structural features into feature vectors required for contrastive learning; the covering manifold prediction and reconstruction module is a prediction and reconstruction head connected after the attention encoding module, which consists of a feedforward network and a transposed convolutional layer, used to project local region features back to the original spatial resolution to generate a single-channel covering manifold reconstruction result.

6. The integrated method for wireless coverage manifold prediction and calibration based on contrastive self-supervision as described in claim 1, characterized in that, In step 3, the specific implementation of the teacher-student comparative self-monitoring method is as follows: The teacher model and the student model adopt the same integrated neural network framework structure; the base station location, terrain height and coverage manifold of the same data sample are simultaneously augmented twice. The augmentation methods include random rotation, random flipping, random cropping and random noise addition, resulting in two different augmented views as input to the teacher model.

7. The integrated method for wireless coverage manifold prediction and calibration based on contrastive self-supervision as described in claim 6, characterized in that, In step 3, the training process employs an attention masking and progressive masking rate strategy: Extract the multi-head attention matrix of the teacher model, calculate the correlation between each local region and the global coverage manifold, and sort them in descending order; A progressive mask rate is adopted: the mask rate gradually increases from 0 to 30% in the early stage of training, from 30% to 70% in the middle stage of training, and gradually increases to 100% full mask in the later stage of training. Based on the current masking rate, select the local regions with the lowest relevance as masking regions, construct a masking matrix to mask the covering manifold of the augmented view, and obtain the masked augmented view as the input of the student model.

8. The integrated method for wireless coverage manifold prediction and calibration based on contrastive self-supervision as described in claim 7, characterized in that, In step 3, the total loss function is constructed as follows: 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. A self-supervised reconstruction loss is constructed based on the reconstruction results of the covered manifold output by the student model and the corresponding original samples, which respectively constrains the reconstruction errors of the masked region and the unmasked region. The total loss function is a weighted sum of the contrastive learning loss and the self-supervised reconstruction loss: in, To compare learning loss, To compensate for losses in self-monitoring and reconstruction, These are the weighting coefficients; 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, To update the weights, and These are the parameters for the student model and the teacher model, respectively.

9. The integrated method for wireless coverage manifold prediction and calibration based on contrastive self-supervision as described in 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 base station location sample, terrain height sample and full mask coverage manifold sample into the model, and output the initial coverage manifold prediction result after one forward inference; Extract the multi-head attention matrix of the last Transformer encoding block, and then... The final attention matrix is ​​obtained by averaging the attention matrices of the individual attention heads. Take the vector after removing the first element from the first row of the matrix as the correlation vector between each local region and the global covering manifold; Select items based on relevance ranking, threshold filtering, or preset quantity filtering, ensuring that the relevance meets the preset criteria. A local area is designated as the key sampling area.

10. The integrated method for wireless coverage manifold prediction and calibration based on contrastive self-supervision as described in claim 1, characterized in that, In step 5, the specific operations for the secondary calibration include: The set of sampling points is determined based on the key sampling area, and a binary sampling position matrix is ​​constructed, where an element of 1 indicates that the corresponding position is a sampling point, and an element of 0 indicates that it is not a sampling point. Based on the real coverage manifold of the test set, the sampling position matrix is ​​multiplied element by element with the coverage manifold of the corresponding region to extract the measured coverage value of the key sampling region, and a sparse measured coverage manifold sample is obtained where only the sampling positions retain valid values ​​and the remaining positions are zero. Base station location samples, terrain height samples, and measured coverage manifold samples are input into the same integrated neural network framework, and the final calibrated wireless coverage manifold is output after secondary inference.

Images