Channel gain map reconstruction method and system based on cross-cluster fusion and spatial neighbor
By using a channel gain map reconstruction method based on cross-cluster fusion and spatial nearest neighbor, regular grid sampling and clustering algorithms are used to divide the region. Combined with Euclidean distance to select k-nearest neighbors and exponential kernel model, the problem of high overhead and information loss in the existing technology is solved, and high-precision and robust channel gain prediction is achieved.
Patent Information
- Application Number
- CN202511666409.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-14
- Publication Date
- 2026-02-10
AI Technical Summary
Existing channel gain map reconstruction methods rely on high-frequency real-time measurements and explicit priors in large-scale MIMO and multi-user concurrent environments, resulting in high overhead and information loss, especially in complex environments where prediction accuracy and robustness are insufficient.
A channel gain map reconstruction method based on cross-cluster fusion and spatial nearest neighbor is adopted. The method divides the region by regular grid sampling and clustering algorithm, selects k nearest neighbors by Euclidean distance, and integrates multi-cluster channel parameters by exponential kernel model to predict channel gain.
It significantly reduces the reliance on real-time channel state information measurement and feedback, improves prediction accuracy and robustness, enhances generalization ability in heterogeneous environments, and reduces signaling overhead.
Smart Images
Figure CN121508690A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of wireless communication, and in particular to a channel gain map reconstruction method and system based on cross-cluster fusion and spatial proximity. BACKGROUND
[0002] In recent years, wireless communication systems still mainly rely on the closed-loop mechanism of "real-time measurement + reporting" to obtain channel state information. In the environment of large-scale MIMO and multi-user concurrency, the terminal needs to periodically return the measurement results, and the frequent signaling not only occupies the air interface resources, but also brings additional time delay and system overhead. Facing the goals of ultra-high capacity, ultra-low latency and massive connection of 6G, the cost and scalability problems of such high-frequency and intensive feedback are increasingly prominent. In order to alleviate the dependence on real-time measurement, the industry has proposed channel knowledge map (CKM) and its core instance - channel gain map (CGM), which can support network planning, optimization and online inference with less online measurement by establishing a mapping between spatial position and channel parameters.
[0003] The existing CGM reconstruction mainly includes four categories: interpolation method, deterministic model, model-driven statistical inference and data-driven method. Although the interpolation method (such as inverse distance weighting, nearest neighbor, Kriging) is simple to implement, the error is easy to amplify in typical scenarios such as strong occlusion, boundary mutation, anisotropy and non-stationarity; the deterministic method electromagnetic simulation (such as ray tracing) can describe the physical propagation, but it is highly dependent on the environment modeling and has high computational cost in large scenes; the data-driven method (such as deep learning) can fit strong nonlinear relationships, but usually requires sufficient and balanced effective samples; the model-driven statistical method is represented by the classic MMSE, and the typical process is to first perform semantic partitioning based on explicit prior (such as LoS / NLoS), then estimate the path loss and shadowing parameters in each sub-region independently, and only select K samples from the same category as the estimated location to participate in the inference during prediction. This process has some deficiencies in engineering, such as excessive dependence on prior knowledge and high cost of prior knowledge acquisition. Secondly, the neighbor selection is limited to the intra-cluster, ignoring the samples outside the cluster but closer in space and more relevant in information, resulting in decreased prediction accuracy.
[0004] Therefore, under the constraints of affordable sampling budget and online complexity, how to simultaneously reduce the dependence on explicit prior (especially LoS / NLoS) and high-density measurement, alleviate the information loss caused by intra-cluster neighbor selection, and improve the robustness and generalization ability in complex and non-stationary environments has become a key pain point for CGM reconstruction. SUMMARY
[0005] The purpose of this invention is to address the shortcomings of existing technologies by using a clustering algorithm to achieve region partitioning and parameter estimation under sparse, regular grid sampling conditions, avoiding the use of expensive prior knowledge for region partitioning. In the prediction stage, cross-cluster information is fused based on spatial proximity, balancing model interpretability with the effective utilization of nearest neighbor information, to achieve high-precision, practical CGM reconstruction under low-overhead conditions.
[0006] To achieve the above objectives, the present invention adopts the following technical solution:
[0007] A channel gain map reconstruction method based on cross-cluster fusion and spatial nearest neighbor includes the following steps:
[0008] S1: Sampling: Sample points using a standard grid and save them;
[0009] S2: Use clustering algorithms to divide the region to better estimate channel parameters;
[0010] S3: After the sampling area is divided, the channel parameter estimation algorithm is used to estimate the channel parameters in each segmented area for the final channel gain prediction.
[0011] S4: After obtaining the channel parameters of each region, the k-nearest neighbor strategy is used to predict the channel gain of the point to be estimated.
[0012] Preferably, in step S4, cross-cluster fusion prediction is used, that is, the global k nearest neighbors are obtained through Euclidean distance, each neighbor retains its original cluster members, and the corresponding channel parameters are retained. .
[0013] Preferably, the specific steps of S4 are as follows:
[0014] In position The predicted channel gain at point is expressed as:
[0015]
[0016] In cross-cluster fusion channel gain prediction, neighborhood It is constructed by selecting k closest sampling points in Euclidean space;
[0017] To integrate information from these different clusters, an exponential kernel is used to model spatial correlations:
[0018]
[0019] In the above formula, Let be a vector representing the distances between data points. For the first Channel parameters for each cluster;
[0020] Shadow component is estimated as:
[0021]
[0022] where is the correlation between the shadow at the target location and the shadow at each neighboring location;
[0023] is the spatial correlation matrix for forming symmetry among k neighbors; is the regularized identity matrix; represents the actual channel gain of k neighbors;
[0024] represents the logarithmic distance from each neighbor to the transmitter;
[0025] The final predicted channel gain is obtained by weighting the predictions of all neighbors, each neighbor using its own channel parameters, thus obtaining the channel gain value for each point to be estimated:
[0026]
[0027] where follows the normalized inverse distance weighting with an exponent of 1, so that closer neighbors obtain greater weights;
[0028] After predicting the channel gain for each estimated point, the corresponding mean square error can be obtained:
[0029] .
[0030] Preferably, the K-means clustering algorithm is used in S2 to divide the area.
[0031] Preferably, the channel parameters in S3 include path loss intercept , path loss exponent , shadow variance , shadow correlation distance and shadow fading parameter .
[0032] Preferably, the path loss intercept and path loss exponent in S3 are estimated using the following formula:
[0033]
[0034] where It is the channel gain vector corresponding to all N sampling points within this sub-region; where It can be represented as ;
[0035] .
[0036] Preferably, the shadow variance Shadow-related distance The estimation methods for fading parameters are as follows:
[0037] After obtaining the estimated path loss parameters, the shading and fading at each data collection location can be obtained by subtracting the path loss:
[0038]
[0039] Based on the above equation, the position is obtained. The residual channel gain of each pair of sampling points at point is denoted as . and And calculate the correlation between these remaining channel gains as follows: This is used to estimate shadow variance and shadow correlation distance;
[0040] set up Where M = N * N;
[0041] Will Redundant and duplicate values are removed, and the size of the newly formed distance set should be less than [the specified value]. And it is represented as ;
[0042] Define a set , The It is a set of distance pairs, and then the correlation function is estimated. Value
[0043]
[0044] The estimation of the shading parameters can be expressed as:
[0045]
[0046] in , indicating that it is estimated At all distances A column vector consisting of logarithmic values;
[0047] , , In the current context of the relevant functions , The set satisfies the calculation The number of data point pairs, estimated by the above formula, is given by the following formula for the multipath fading variance:
[0048] .
[0049] This application also provides a channel gain map reconstruction system based on cross-cluster fusion and spatial proximity, characterized in that: the system adopts any one of the methods of claims 1-8 as its basic algorithm.
[0050] Preferably, the system can be deployed on servers, base stations, or user terminal equipment in a cellular network.
[0051] Compared with the prior art, this application has the following beneficial effects:
[0052] 1. This invention proposes a channel gain prediction mechanism based on multi-cluster parameter fusion and spatial proximity. It does not rely on prior environmental labels such as NLoS / LoS. It can achieve high-precision CGM reconstruction based only on the spatial distribution of limited sampling points and channel measurement values. This significantly reduces the system's dependence on real-time channel state information measurement and feedback, and effectively reduces signaling overhead.
[0053] 2. By fusing channel parameters from multiple clusters and selecting reference points based on spatial proximity, this invention can make fuller use of channel spatial correlation, overcome the boundary effects and information loss in traditional homogeneous region channel gain prediction methods, significantly improve prediction accuracy, and enhance the robustness and generalization ability of the algorithm in heterogeneous environments. Attached Figure Description
[0054] Figure 1 This paper compares the performance of cross-cluster fusion channel gain prediction and homogeneous region channel gain prediction algorithms in one embodiment of this application.
[0055] Figure 2 It is a real CGM;
[0056] Figure 3 This is the CGM for cross-cluster fusion channel gain prediction in one embodiment of this application. Detailed Implementation
[0057] The present invention will be further described in detail below with reference to specific embodiments.
[0058] To overcome the problems of existing CGM reconstruction methods, such as reliance on high-cost field measurements and explicit priors (e.g., LoS / NLoS labels), information loss due to limited intra-cluster neighbor selection, high online prediction complexity, and extremely large sample size requirements, this invention aims to propose a low-overhead, high-precision CGM reconstruction method and system that can be deployed under sparse, regular grid sampling conditions. This method aims to achieve multi-cluster parameter fusion and adaptive neighbor selection based on spatial proximity, significantly reducing the mean square error of CGM point prediction, improving prediction stability and generalization ability in heterogeneous scenarios without increasing additional labeling and measurement inputs, and keeping the online inference complexity within an engineering-acceptable range. This provides directly applicable, high-confidence channel gain information for practical network planning and operation.
[0059] Specifically, this application provides a channel gain map reconstruction method based on cross-cluster fusion and spatial nearest neighbor, comprising the following steps:
[0060] S1: Collect and save the sampling points using a regular grid (with different d intervals);
[0061] Specifically, data is collected at intervals d both horizontally and vertically, excluding building sections. All valid sampling points are saved. Each sampling point can be represented as follows: Let this set of sampling points be denoted as There are a total of N sampling points. The first two components represent the horizontal and vertical coordinates of the sampling point, respectively, and the last component is the channel gain corresponding to the sampling point.
[0062] S2: In order to better capture the wireless propagation environment of effective sampling points, a clustering algorithm is used to divide the region to better estimate channel parameters.
[0063] In one implementation, the K-means clustering algorithm is used to divide the region.
[0064] S3: After the sampling area is divided, the channel parameter estimation algorithm is used to estimate the channel parameters in each segmented area for the final channel gain prediction.
[0065] In one embodiment, the channel parameters include path loss intercept. Path loss index Shadow variance Shadow-related distance and shadow fading parameters .
[0066] Due to shadowing correlation, joint maximum likelihood estimation of all channel parameters is complex and challenging. Therefore, it is necessary to estimate path loss parameters and shadowing fading parameters separately.
[0067] First, estimate the path loss intercept. and path loss index And given by the following equation:
[0068]
[0069] in It is the channel gain vector corresponding to all N sampling points within this sub-region. It can be represented as .
[0070] .
[0071] Next, the shadowing and fading parameters are estimated.
[0072] Specifically, after obtaining the estimated path loss parameters, the shading and fading at each data collection location can be obtained by subtracting the path loss:
[0073]
[0074] Based on the above equation, the position can be obtained. The residual channel gain of each pair of sampling points at point is denoted as . and And calculate the correlation between these remaining channel gains as follows: It can be used to estimate shadow variance and shadow correlation distance.
[0075] set up Where M = N*N. To better obtain the correlation values associated with each distance, Redundant and duplicate values are removed, and the size of the newly formed distance set should be less than [the specified value]. And it is represented as .
[0076] Define a set , The aforementioned It is a set of distance pairs. It can be considered a very small number, in one embodiment, the... Set it to 0.1. Then estimate the correlation function. Value
[0077]
[0078] To reduce computational complexity and make the results meaningful, the negative correlation function values obtained from the calculation are discarded, and the computational complexity is also reduced. The scale, when It needs to meet the following requirements. Therefore, the estimation of the shadow parameters can be expressed as:
[0079]
[0080] in , indicating that it is estimated At all distances A column vector composed of logarithmic values.
[0081] , . In the current context of the relevant functions . The set satisfies the calculation The number of data point pairs. The multipath fading variance estimated from the above formula is given by the following formula:
[0082]
[0083] S4: After obtaining the channel parameters of each region, the k-nearest neighbor strategy is used to predict the channel gain of the point to be estimated.
[0084] This application provides two k-nearest neighbor strategies to predict the channel gain of the point to be estimated.
[0085] The first method: Homogeneous region prediction
[0086] The points to be estimated are clustered simultaneously, so the sampling points and the points to be estimated can obtain the corresponding categories at the same time. The k nearest neighbor points to be predicted of the point to be estimated must belong to the same category as this point to be estimated. This can be understood as channel gain prediction in homogeneous regions.
[0087] The second type: Cross-cluster fusion prediction
[0088] The global k-nearest neighbor obtained through Euclidean distance is called cross-cluster fusion channel gain prediction. Each neighbor in CF-CGP retains its original cluster members and corresponding channel parameters. They are used separately in the prediction process.
[0089] This method allows for multi-region parameter fusion, enhancing robustness in heterogeneous environments. Once the cluster parameters are estimated, channel gain prediction focuses on the shadow component. The above is because large-scale path loss deterministically depends on the target location. The distance between the base station and the location, while small-scale fading is treated as random noise. The predicted channel gain at point is expressed as:
[0090]
[0091] In cross-cluster fusion channel gain prediction, neighborhood It is constructed by selecting k closest sampling points in Euclidean space.
[0092] While spatial proximity might indicate similar propagation conditions, this is not necessarily the case due to the heterogeneity of real-world environments, where cluster boundaries may not correspond to geometric distances. Therefore, the k nearest neighbors may span multiple clusters. In one implementation, to integrate information from these different clusters, an exponential kernel is used to model spatial correlations:
[0093]
[0094] In the above formula, Let be a vector representing the distances between data points. For the first Channel parameters for each cluster.
[0095] Shadow component The estimate is:
[0096]
[0097] in Used to capture target location The correlation between the shadow at a given location and the shadow at each of its adjacent locations;
[0098] , used to form a symmetric spatial correlation matrix among k neighbors; It is a regularized identity matrix; This represents the actual channel gain of the k neighbors;
[0099] This represents the logarithmic distance from each neighbor to the transmitter.
[0100] The final predicted channel gain is obtained by weighting the predictions of all neighbors, with each neighbor using its own channel parameters, thus yielding the channel gain value for each point to be estimated:
[0101]
[0102] in It follows a normalized inverse distance weighting with an exponent of 1, meaning each neighbor is assigned a weight that is inversely proportional to its distance from the query, and the weights are normalized to sum to 1, so closer neighbors get a larger weight.
[0103] After predicting the channel gain at each estimated point, the corresponding mean square error can be obtained:
[0104] .
[0105] Based on the above channel prediction method, this application also provides a CGM prediction system based on multi-cluster fusion and spatial proximity, wherein the system adopts the prediction algorithm described above.
[0106] The system can be deployed on servers, base stations, or user terminal equipment within cellular networks. It is particularly suitable for scenarios such as network planning and optimization in 5G-Advanced or 6G networks, millimeter-wave beam management, UAV communication path planning, and link budget estimation in large-scale IoT applications. Based on limited channel sampling data, it can quickly reconstruct an accurate channel gain map covering the entire target area, providing underlying channel information support for various wireless applications.
[0107] In summary, the purpose of this application is to provide a CGM reconstruction method based on multi-cluster parameter fusion and spatial proximity. This method achieves region division and multi-region parameter estimation through clustering algorithms. During prediction, it breaks through the traditional co-cluster limitation, selects neighboring points based on spatial Euclidean distance, and fuses multi-cluster channel information for channel gain prediction. This effectively reduces the dependence on a large number of sampling points and prior environmental information, and significantly improves the prediction accuracy and robustness of the channel map under limited sampling.
[0108] In addition, please see Figures 1-3 The methods described above have been extensively tested in MATLAB, among which... Figure 1 A comparative analysis of MSE (Mean Sequence Equation) for homogeneous region channel gain prediction and cross-cluster fusion channel gain prediction in different sampling intervals is presented. To ensure fairness and control computational complexity, the number of nearest neighbors for both methods is fixed at K=3.
[0109] For K=3, MSE continues to decline, but the marginal improvement becomes smaller; therefore, K=3 provides a representative trade-off between reporting accuracy and computational cost.
[0110] As the sampling interval decreases, the number of effective samples increases, which improves the prediction accuracy of both methods and produces a lower MSE.
[0111] In addition to this overall trend, cross-cluster fusion channel gain prediction consistently outperforms homogeneous region channel gain prediction at any sampling interval, with the performance gap widening at finer sampling. This suggests that information fusion from heterogeneous clusters becomes increasingly effective when more samples are available near each query location.
[0112] Using nlos / los for partitioning leverages prior knowledge for region division. It can be seen that even without prior knowledge, the k-means clustering algorithm can still reconstruct a more accurate CGM, with an MSE of approximately 0.40 dB. The results show that the channel gain prediction model described in this application exhibits significant advantages and is more suitable for reconstructing high-precision CGMs than existing methods for channel gain prediction in homogeneous regions. The original CGM obtained from ray tracing is as follows: Figure 2 As shown, the CGM reconstructed after cross-cluster fusion channel gain prediction is as follows: Figure 3 As shown, the reconstructed CGM has basically restored the reality.
Claims
1. A channel gain map reconstruction method based on cross-cluster fusion and spatial nearest neighbor, characterized in that: Includes the following steps: S1: Sampling: Sample points using a standard grid and save them; S2: Use clustering algorithms to divide the region to better estimate channel parameters; S3: After the sampling area is divided, the channel parameter estimation algorithm is used to estimate the channel parameters in each segmented area for the final channel gain prediction. S4: After obtaining the channel parameters of each region, the k-nearest neighbor strategy is used to predict the channel gain of the point to be estimated.
2. The channel gain map reconstruction method based on cross-cluster fusion and spatial nearest neighbor as described in claim 1, characterized in that: In S4, cross-cluster fusion prediction is used, which is based on the global k nearest neighbors obtained through Euclidean distance. Each neighbor retains its original cluster members and corresponding channel parameters. .
3. The channel gain map reconstruction method based on cross-cluster fusion and spatial nearest neighbor as described in claim 1, characterized in that: The specific steps of S4 are as follows: In position The predicted channel gain at point is expressed as: In cross-cluster fusion channel gain prediction, neighborhood It is constructed by selecting k closest sampling points in Euclidean space; To integrate information from these different clusters, an exponential kernel is used to model spatial correlations: In the above formula, Let be a vector representing the distances between data points. For the first Channel parameters for each cluster; Shadow component The estimate is: in Used to capture target location The correlation between the shadow at a given location and the shadow at each of its adjacent locations; Used to form a symmetric spatial correlation matrix among k neighbors; It is a regularized identity matrix; This represents the actual channel gain of the k neighbors; This represents the logarithmic distance from each neighbor to the transmitter; The final predicted channel gain is obtained by weighting the predictions of all neighbors, with each neighbor using its own channel parameters, thus yielding the channel gain value for each point to be estimated: in It follows a normalized inverse distance weighting with an exponent of 1, so closer neighbors get a larger weight; After predicting the channel gain at each estimated point, the corresponding mean square error can be obtained: 。 4. The channel gain map reconstruction method based on cross-cluster fusion and spatial nearest neighbor as described in claim 1, characterized in that: The region is divided using the K-means clustering algorithm in S2.
5. The channel gain map reconstruction method based on cross-cluster fusion and spatial nearest neighbor as described in claim 1, characterized in that: The channel parameters in S3 include path loss intercept. Path loss index Shadow variance Shadow-related distance and shadow fading parameters .
6. The channel gain map reconstruction method based on cross-cluster fusion and spatial nearest neighbor as described in claim 5, characterized in that: The path loss intercept in S3 and path loss index The following formula is used for estimation: in It is the channel gain vector corresponding to all N sampling points within this sub-region; where It can be represented as ; 。 7. The channel gain map reconstruction method based on cross-cluster fusion and spatial nearest neighbor as described in claim 6, characterized in that: The shadow variance Shadow-related distance The estimation methods for fading parameters are as follows: After obtaining the estimated path loss parameters, the shading and fading at each data collection location can be obtained by subtracting the path loss: Based on the above equation, the position is obtained. The residual channel gain of each pair of sampling points at point is denoted as . and ; And calculate the correlation between these remaining channel gains. This is used to estimate shadow variance and shadow correlation distance; set up Where M = N * N; Will Redundant and duplicate values are removed, and the size of the newly formed distance set should be less than [the specified value]. And it is represented as ; Define a set , The It is a set of distance pairs, and then the correlation function is estimated. Value The estimation of the shading parameters can be expressed as: in , indicating that it is estimated At all distances A column vector consisting of logarithmic values; , , In the current context of the relevant functions , The set satisfies the calculation The number of data point pairs, estimated by the above formula, is given by the following formula for the multipath fading variance: 。 8. A channel gain map reconstruction system based on cross-cluster fusion and spatial nearest neighbor, characterized in that: The system adopts any one of the methods in claims 1-7 as its basic algorithm.
9. A channel gain map reconstruction system based on cross-cluster fusion and spatial nearest neighbor as described in claim 8, characterized in that: The system can be deployed on servers, base stations, or user terminal equipment in a cellular network.