Channel gain map reconstruction method based on k neighbor refined neural network
By combining global and local neighborhood information through a k-nearest neighbor refined neural network, the problems of high-density sampling and explicit prior dependence in CGM reconstruction are solved, achieving high-precision channel gain prediction, reducing costs and improving prediction stability. It is suitable for coverage planning and environmental awareness in cellular networks.
Patent Information
- Application Number
- CN202511825035.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-05
- Publication Date
- 2026-02-27
AI Technical Summary
Existing CGM reconstruction techniques rely heavily on high-density sampling measurements and explicit priors in regular grid sampling scenarios, and online inference is complex, making it difficult to achieve high-precision channel gain prediction without increasing costs.
A method based on k-nearest neighbor refined neural networks is adopted. By combining preliminary prediction by fully connected neural networks and radial basis function support vector regression, channel gain is reconstructed using local and global neighborhood information, reducing the dependence on high-density sampling and explicit priors. Combined with clustering algorithms and nearest neighbor selection strategies, high-precision prediction of channel gain is achieved.
High-precision channel gain reconstruction is achieved under sparse regular grid sampling, which reduces measurement and maintenance costs, improves prediction stability and generalization ability in complex scenarios, and is suitable for coverage planning and environmental perception of cellular networks.
Smart Images
Figure CN121585295A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of wireless communication technology, and in particular to a method for reconstructing a channel gain map based on a k-nearest neighbor refined neural network. Background Technology
[0002] In recent years, wireless communication systems have placed higher demands on accurate and low-latency channel information. The traditional "measurement-feedback" paradigm introduces significant latency overhead in massive MIMO and multi-user intensive scenarios, making it difficult to meet the 6G goals of high capacity, low latency, and massive connectivity. To reduce reliance on online CSI, concepts such as the Channel Knowledge Map (CKM) framework have emerged, with the Channel Gain Map (CGM) as a key component. The CGM provides channel gain knowledge (e.g., including geographical coordinates and channel gain information) for any location within a cell. Assuming the CGM is static, it can be used for base station resource allocation and also facilitates the design, optimization, and management of wireless communication systems.
[0003] Existing CGM reconstruction approaches mainly include: deterministic electromagnetic simulation (such as ray tracing), field measurements, traditional interpolation methods, and model-based methods. Ray tracing is computationally expensive and highly dependent on environmental modeling; large-scale field measurements are difficult to sustain due to manpower and other costs. Traditional interpolation methods, while simple to implement, suffer from significantly amplified errors in typical scenarios such as significant occlusion and abrupt boundary changes. Model-based methods require costly explicit prior knowledge to ensure the stability of parameter learning, but the acquisition and maintenance of such priors are costly and have poor transferability. Data-driven methods, such as deep learning, can predict channel gain, but typically require a large number of samples.
[0004] Therefore, there is a current need for a CGM reconstruction technique that operates under regular grid sampling conditions (sparse, not relying on too many samples), has low dependence on explicit priors, and offers controllable online prediction costs, providing a scalable and feasible solution for engineering deployment. Without increasing measurement and manual annotation costs, this technique leverages neighborhood spatial context, balancing global trend modeling with local refinement, thereby significantly reducing mean squared error and improving robustness and generalization ability in complex scenarios. Summary of the Invention
[0005] The purpose of this application is to overcome the problems of existing CGM reconstruction methods' reliance on high-density sampling measurements and explicit priors (such as LosS / NLoS labels) in regular grid sampling scenarios, as well as the high complexity of online inference. This application provides a low-overhead and high-precision CGM reconstruction method under sparse regular grid sampling conditions. The goal is to explicitly utilize neighborhood information based on spatial distance without increasing the workload of sampling measurements and manual annotation, and to combine global trends with local refinement to achieve accurate prediction of channel gain at any query location. This significantly reduces the reconstruction error of CGM (i.e., the mean square error of channel gain prediction at the query location, MSE), and improves the prediction stability and generalization ability in complex propagation scenarios such as boundaries and occlusion.
[0006] To achieve the above objectives, the present invention adopts the following technical solution:
[0007] A channel gain map reconstruction method based on a k-nearest neighbor refined neural network includes the following steps:
[0008] S1: Data Acquisition: Obtain a limited set of measurement samples by sampling according to a regular grid pattern in the target area. ;
[0009] S2: Constructing a Fully Connected Neural Network (FCN) , as the first stage of the neural network;
[0010] S3: Divide the finite measurement sample set D into multiple clusters using a clustering algorithm, obtain the cluster center of each cluster, calculate the Euclidean distance between the location of the query point and the cluster center of each cluster, and assign the query point to different clusters;
[0011] S4: For each query position Use at least two nearest neighbor selection strategies.
[0012] S5: Construction of the k-Nearest Neighbor Refinement Network: The initial prediction from the first-stage neural network's FCN is used as the "anchor label" for the query point. Different local regression training sets are obtained under different nearest neighbor strategies. The initial prediction information of the query point is then used to train the radial basis kernel support vector regression (SVR) again with the different local regression training sets. The key hyperparameters of SVR are determined by the small grid and validation set to minimize MSE. After training, the query position is used again... As input, the trained local SVR is invoked to obtain the final prediction: This output is directly used as the refined channel gain estimate.
[0013] Preferably, in step S1, after collecting samples horizontally at intervals d within the CGM scene, longitudinally at intervals d are then collected, skipping building sections and not collecting samples. This process continues until the entire target area is collected, thus creating a set of all collected finite measurement samples. This serves as the training set for the first stage of neural network training.
[0014] Preferably, the network structure in S2 adopts a multi-layer fully connected structure, with the hidden layer activation function being ReLU and the output layer activation function being linear.
[0015] Preferably, the first-stage neural network constructed in S2 has an input layer of two neurons, a total of nine hidden layers, and an output layer of one neuron. Its optimization algorithm is set to scale the conjugate gradient. Iterative minimization.
[0016] Preferably, in step S2, the training set is further divided into training / validation / test in a ratio of 8:1:1. After each epoch of training, the MSE is calculated on the validation set. If the validation MSE does not improve for n consecutive rounds (n≥10), training is stopped and the parameters are rolled back to the round with the optimal validation MSE. The combination with the smallest validation MSE is selected as the final FCN structure and hyperparameters.
[0017] Preferably, the clustering algorithm is one of K-means, Gmm, or Pam.
[0018] Preferably, there are two strategies in S4. Strategy 1: After obtaining the labels of each query point in step S3, find the k nearest neighbors with the same cluster labels outside each query point to obtain the intra-cluster neighborhood. ;
[0019] Strategy 2: For the query point, directly select the k nearest neighbors in the sample set D according to the Euclidean distance to obtain the global neighborhood. Therefore, the neighborhood may contain sampling points from multiple clusters.
[0020] Preferably, the preliminary prediction information includes its coordinates and the preliminary predicted channel gain.
[0021] Compared with the prior art, this application has the following beneficial effects:
[0022] Compared with existing technologies, the present invention has the following advantages: 1. By adopting a k-nearest neighbor refined network structure, CGM can achieve high-precision reconstruction under sparse regular grid sampling.
[0023] 2. The method provided in this application does not rely on costly prior knowledge such as LoS / NLoS throughout the entire process; since it allows cross-cluster neighbor selection, the prediction stability is significantly better than the intra-cluster approach, and it makes better use of local spatial continuity.
[0024] 3. In this application, the FCN in Phase 1 is trained offline only once; in Phase 2, each query point only requires one nearest neighbor search and fitting with a small sample of K+1 data points and SVR. The current implementation uses full distance sorting to select the k nearest data points, with a time complexity of O(n). SVR training / inference is performed on small samples, making it suitable for online engineering applications.
[0025] 4. The construction method provided in this application reduces the reliance on high-density field measurements and explicit tags while maintaining or improving prediction accuracy; thereby reducing deployment and maintenance costs and facilitating its use in daily operation and maintenance scenarios such as coverage planning and environmental perception in cellular networks (including 5G / 6G). Attached Figure Description
[0026] Figure 1 This is a diagram of the k-nearest neighbor refined network structure in one embodiment of this application;
[0027] Figure 2 This is a performance graph of the k-nearest neighbor refinement network in one embodiment of this application;
[0028] Figure 3 The original CGM was obtained through ray tracing;
[0029] Figure 4 This is the CGM that was reconstructed after prediction. Detailed Implementation
[0030] The present invention will be further described in detail below with reference to specific embodiments.
[0031] This application provides a channel gain map reconstruction method based on a k-nearest neighbor refined neural network, which is oriented towards two-dimensional planar location. Perform channel gain prediction to reconstruct the CGM, with the sampled dataset as input. ,in For the coordinates of the sampling points, This corresponds to the measured channel gain (in dB). Targeting the query location set. Output the channel gain for each query point. Channel gain information can be queried at any location within the cell via CGM.
[0032] The specific steps of the reconstruction method are as follows:
[0033] S1: Data Acquisition: Obtain a limited set of measurement samples by sampling according to a regular grid pattern in the target area. .
[0034] Specifically, in one implementation, after collecting samples horizontally at intervals d within the CGM scene, longitudinally at intervals d are then collected, skipping building sections and not collecting samples. This process continues until the entire target area is collected, thus creating a set of all collected finite measurement samples. This serves as the training set for the first stage of neural network training.
[0035] In one embodiment, the interval d = 5.
[0036] S2: Constructing a Fully Connected Neural Network (FCN) This serves as the first stage of the neural network.
[0037] Specifically, input location The channel gain predicted in the first stage is obtained. .
[0038] The network structure in this application adopts a multi-layer fully connected structure, with ReLU activation function in the hidden layer and linear activation function in the output layer.
[0039] The input layer consists of two neurons, and there are nine hidden layers with the following number of neurons: a, b, c, d, e, d, c, b, a. The output layer consists of one neuron. In one embodiment, the number of neurons is 64, 128, 256, 512, 1024, 512, 256, 128, 64. The optimization algorithm is set to scale the conjugate gradient. Iterative minimization. Scaling conjugate gradients belong to the second-order approximation conjugate gradient method, possessing the characteristics of adaptive step size and no need for line search, making it suitable for regression in medium-sized fully connected networks.
[0040] This application also employs a strategy of early training termination to avoid overfitting and ineffective iterations. Specifically, in one embodiment, the training set is divided into training / validation / test sets in an 8:1:1 ratio. After each epoch of training, the MSE is calculated on the validation set. If the validation MSE does not improve for n consecutive epochs (n≥10), training is stopped and the parameters are reverted to the epoch with the optimal validation MSE. The combination with the smallest validation MSE is selected as the final FCN structure and hyperparameters.
[0041] S3: Divide the finite measurement sample set D into multiple clusters using classic clustering algorithms such as K-means, Gmm, and Pam, and obtain the cluster center of each cluster. Calculate the Euclidean distance between the location of the query point and the cluster center of each cluster, and assign the query point to different clusters. Thus, the query point has a cluster label.
[0042] S4: For each query position There are two nearest neighbor selection strategies.
[0043] Strategy 1: Obtain the labels of each query point through step S3, find the k nearest neighbors with the same cluster labels outside each query point, and obtain the intra-cluster neighborhood. .
[0044] Strategy 2: For the query point, directly select the k nearest neighbors in the sample set D according to the Euclidean distance to obtain the global neighborhood. Therefore, the neighborhood may contain sampling points from multiple clusters.
[0045] S5: Construction of k-nearest neighbor refined network:
[0046] Please see Figure 1 The initial prediction of the first-stage neural network FCN is used as the "anchor label" of the query point. Different local regression training sets are obtained under two nearest neighbor strategies. The initial prediction information of the query point and the different local regression training sets are used to train the radial basis kernel support vector regression (SVR). The key hyperparameters of SVR are determined by the small grid and the validation set to minimize MSE. The initial prediction information includes its coordinates and the initial predicted channel gain.
[0047] After training is complete, query the location again. As input, the trained local SVR is invoked to obtain the final prediction: This output is directly used as the refined channel gain estimate, and this two-stage neural network structure is named the k-nearest neighbor refinement network.
[0048] The methods described above in this application have been tested in MATLAB. Please refer to [link / reference]. Figure 2 and Figure 3 , Figure 2 The results are shown for the k-nearest neighbor (kNN) refinement network, neighboring points, and the kNN strategy (taking d=5 as an example). It can be seen that, with the kNN refinement network as the baseline, the global kNN strategy achieves a lower mean square error (MSE) for better CGM reconstruction. As k increases, the MSE of intra-cluster nearest neighbor strategies based on different partitioning algorithms significantly increases on the kNN refinement network. The results demonstrate that the channel gain prediction model described in this patent exhibits significant advantages and is more suitable for reconstructing high-precision CGMs than existing methods. The original CGM obtained from ray tracing is shown below. Figure 3 As shown, the CGM reconstructed after prediction is as follows: Figure 4 As shown, the reconstructed CGM has basically restored the real channel.
Claims
1. A channel gain map reconstruction method based on a k-nearest neighbor refined neural network, characterized in that: Includes the following steps: S1: Data Acquisition: Obtain a limited set of measurement samples by sampling according to a regular grid pattern in the target area. ; S2: Constructing a Fully Connected Neural Network (FCN) , as the first stage of the neural network; S3: Divide the finite measurement sample set D into multiple clusters using a clustering algorithm, obtain the cluster center of each cluster, calculate the Euclidean distance between the location of the query point and the cluster center of each cluster, and assign the query point to different clusters; S4: For each query position Use at least two nearest neighbor selection strategies. S5: Construction of the k-Nearest Neighbor Refinement Network: The initial prediction from the first-stage neural network's FCN is used as the "anchor label" for the query point. Different local regression training sets are obtained under different nearest neighbor strategies. The initial prediction information of the query point is then used to retrain the radial basis kernel support vector regression (SVR) with the different local regression training sets. The key hyperparameters of SVR are determined by the small grid and validation set to minimize MSE. After training, the query position is used again... As input, the trained local SVR is invoked to obtain the final prediction: This output is directly used as the refined channel gain estimate.
2. The channel gain map reconstruction method based on k-nearest neighbor refined neural network according to claim 1, characterized in that: In step S1, after collecting samples horizontally at intervals d within the CGM scene, longitudinally at intervals d are then collected, skipping building sections and not collecting samples. This process continues until the entire target area is collected, thus creating a set of all collected finite measurement samples. This serves as the training set for the first stage of neural network training.
3. The channel gain map reconstruction method based on k-nearest neighbor refined neural network according to claim 1, characterized in that: The network structure in S2 adopts a multi-layer fully connected structure, with ReLU activation function in the hidden layer and linear activation function in the output layer.
4. The channel gain map reconstruction method based on k-nearest neighbor refined neural network according to claim 3, characterized in that: The first-stage neural network constructed in S2 has an input layer of two neurons, a total of nine hidden layers, and an output layer of one neuron. Its optimization algorithm is set to scale the conjugate gradient. Iterative minimization.
5. The channel gain map reconstruction method based on k-nearest neighbor refined neural network according to claim 4, characterized in that: In S2, the training set is divided into training / validation / test in a ratio of 8:1:
1. After each epoch of training, the MSE is calculated on the validation set. If the validation MSE does not improve for n consecutive (n≥10) rounds, training is stopped and the parameters are rolled back to the round with the optimal validation MSE. The combination with the smallest verification MSE is selected as the final FCN structure and hyperparameters.
6. The channel gain map reconstruction method based on k-nearest neighbor refined neural network according to claim 1, characterized in that: The clustering algorithm is one of K-means, Gmm, or Pam.
7. The channel gain map reconstruction method based on k-nearest neighbor refined neural network according to claim 1, characterized in that: There are two strategies in S4. Strategy 1: After obtaining the labels of each query point in step S3, find the k nearest neighbors with the same cluster labels outside each query point to obtain the intra-cluster neighborhood. ; Strategy 2: For the query point, directly select the k nearest neighbors in the sample set D according to the Euclidean distance to obtain the global neighborhood. Therefore, the neighborhood may contain sampling points from multiple clusters.
8. The channel gain map reconstruction method based on k-nearest neighbor refined neural network according to claim 1, characterized in that: The preliminary prediction information includes its coordinates and the preliminary predicted channel gain.