Fast communication method based on wireless channel knowledge map construction

Abstract

The application discloses a fast communication method based on a wireless channel knowledge map, and relates to the field of the sixth generation communication technology. Channel parameter matrices of users in communication are collected and processed to generate standard channel parameter matrices of each user, and an adjacency matrix is constructed based on the spatial distance of the users. A graph auto-encoder is used to aggregate features and capture relevance of the standard channel parameter matrices, to generate an encoded feature matrix and map it to a reconstructed adjacency matrix to verify whether to output. A partition clustering algorithm based on community discovery is used to iteratively merge the encoded feature vectors of each user, to generate a final community set and calculate a centroid feature matrix. A weight prediction network is used to expand and map a target position, and the centroid feature matrix is spliced, a weight coefficient vector is obtained through a multi-head attention mechanism and expansion mapping, and the channel parameter matrices of each user are weighted and summed to obtain the channel parameter matrix of the target position, so that the channel knowledge map is accurately constructed and used to establish the communication process of the base station and the user.

Description

Technical Field

[0001] This invention relates to the field of sixth-generation communication technology, and more specifically to a fast communication method based on the construction of a wireless channel knowledge map. Background Technology

[0002] In wireless communication, communication between users and base stations requires prior knowledge of the channel parameter matrix between them. Traditional channel estimation methods estimate the channel parameter matrix by inserting a certain number of known pilot symbols before each data transmission and then constructing a reliable communication process based on the estimation results to transmit data. However, traditional channel estimation methods result in low data transmission efficiency because pilot symbols occupy communication resources.

[0003] As an emerging communication-aided database, the channel knowledge map can accurately depict the distribution patterns of the channel parameter matrix and the map environment, providing data support for resource allocation, link optimization, and interference coordination. By constructing the channel knowledge map, the channel parameter matrix can be associated with spatial location. After obtaining the user's location, the channel parameter matrix can be obtained quickly and reliably by directly querying the channel knowledge map, thus achieving efficient communication. It is a key foundation for promoting the intelligent and efficient development of wireless communication.

[0004] The accuracy of wireless communication depends on the precise construction of the channel knowledge map. Existing methods for constructing channel knowledge maps suffer from problems such as insufficient capture of the aggregation and correlation of channel features, poor accuracy of partition clustering, and poor adaptability of interpolation prediction when facing complex multi-user scenarios. As a result, the constructed channel knowledge map cannot accurately reflect the actual channel situation, which affects the performance of the wireless communication process. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention proposes a fast communication method based on the construction of a wireless channel knowledge map, thereby achieving a fast communication method based on a channel knowledge map constructed with precise partitioning.

[0006] The technical solution to achieve the objective of this invention is as follows:

[0007] A fast communication method based on wireless channel knowledge maps includes the following steps:

[0008] from In the communication process of the first user, the location and channel parameter matrix are obtained, and for the first user... Channel parameter matrix for each user Perform path completion and normalize to the first... Standard channel parameter matrix for each user ,use An adjacency matrix is ​​constructed based on the spatial distance relationships between individual users. ;

[0009] Using a graph autoencoder, for The standard channel parameter matrix of each user is used for feature aggregation and correlation capture to generate a coded feature matrix. Based on the feedback decoder, the encoded feature matrix is... Mapping to reconstruct the adjacency matrix And based on reconstruction loss Decide whether to output the encoded feature matrix ;

[0010] A community-based partitioning clustering algorithm is adopted, based on the encoded feature matrix. Construct a weighted graph and treat each user as a community. Evaluate the similarity between communities based on modularity and iteratively merge them until merging is no longer possible, thus obtaining the final community set. Then, the centroid feature vector of each endgame community is calculated to generate the centroid feature matrix. ,in, For the first An endgame community. The final total number of communities;

[0011] Using a weighted prediction network to determine the target location After dimension expansion mapping and centroid feature matrix The weight coefficient vector is obtained by concatenation, through multi-head attention mechanism and dimension expansion mapping. ,Will The channel parameter matrix of each user is based on the weighted coefficient vector. The target position is obtained by weighting and summing the corresponding weight coefficients. Channel parameter matrix Construct a channel knowledge map;

[0012] Receive communication requests and new user locations from new users. The nearest neighbor algorithm is used to match the channel parameter matrix of new users from the channel knowledge map. To establish a communication process with new users.

[0013] Furthermore, based on the downlink channel model, the first Channel matrix between users and base stations The determination depends on the first Channel parameters for each user, including The arrival pitch angle, arrival azimuth angle, departure pitch angle, departure azimuth angle, gain, and propagation delay of the path. For the first The number of paths for each user, the first The channel parameters of the first user are obtained through a pilot-based channel estimation algorithm, and then organized to construct the first... Channel parameter matrix for each user .

[0014] Furthermore, regarding Each user performs path completion separately to determine... Maximum number of paths per user Each path has a number of paths less than the maximum number of paths. The user's channel parameter matrix is ​​padded with a vector of all zeros so that the number of rows is equal to the maximum number of paths. Consistent, calculate separately The mean and standard deviation of the arrival pitch angle, arrival azimuth angle, departure pitch angle, departure azimuth angle, gain, and propagation delay for each user are calculated and normalized using Z-Score to obtain... A standard channel parameter matrix is ​​used for each user, treating each user as a node. The reciprocal of the spatial distance between two users is calculated and used as an edge connecting the two nodes. A node graph is constructed and stored in matrix form. Edges between users are used to construct an adjacency matrix. Adjacency matrix The Line 1 List For the first The user and the first The reciprocal of the spatial distance between users.

[0015] Furthermore, the graph autoencoder includes an encoder and a feedback decoder;

[0016] The encoder uses graph convolution based on the adjacency matrix. For the Standard channel parameter matrix for each user Feature aggregation is performed to obtain the first Embedded vectors of individual users By capturing the nonlocal correlation of embedding vectors from different users using Gaussian radial basis functions, an encoded feature matrix is ​​generated. ;

[0017] The feedback decoder uses a multilayer perceptron to encode the feature matrix. Mapping to reconstruct the adjacency matrix Calculate the reconstructed adjacency matrix Adjacency Matrix Reconstruction loss The result is compared with the loss threshold; if it is less than or equal to the loss threshold, the encoded feature matrix is ​​output. If the loss exceeds the threshold, the encoder parameters are updated based on gradient descent and the encoded feature matrix is ​​regenerated. Among them, reconstruction loss Equivalent to reconstructing the adjacency matrix Adjacency Matrix Mean square error Add based on the encoding feature matrix The regularization term is calculated.

[0018] Furthermore, the encoder generates an encoded feature matrix. This includes the following steps:

[0019] Adjacency matrix Set all diagonal elements to 1 to generate a self-connected adjacency matrix. Calculate the self-connected adjacency matrix degree matrix ;

[0020] Using the information transfer mechanism of graph convolution to the first Standard channel parameter matrix for each user conduct Secondary feature aggregation generates the th... Embedded vectors of individual users , among which, the The input for the second feature aggregation is the first... The user The aggregation channel parameter matrix of the second order. The output is the first... The user The aggregation channel parameter matrix of the second order. , And the aggregated channel parameter matrix of the 0th iteration. and the The aggregation channel parameter matrix of the second order. The first Standard channel parameter matrix for each user and embedded vector ;

[0021] The Gaussian radial basis function is used to calculate the first... The user and the first Encoded response between users , will encode the response Arranged in the th Line 1 Columns, generate encoded feature matrix .

[0022] Furthermore, the community-based partitioning clustering algorithm includes the following steps:

[0023] The first The user and the Each user is considered a vertex, and the encoded feature matrix is ​​used. The Middle Encoded feature vector of each user With the Encoded feature vector of each user cosine similarity, constructing the first The user and the first Weight edge between users ;

[0024] Based on total number of users Set dynamic weight threshold ,right Prune the weighted edges between users, retaining those greater than the dynamic weight threshold. Construct a weighted graph using the weighted edges;

[0025] Treating each user in the weighted graph as an independent community, and defining communities with weighted edges as neighboring communities, for the i-th a community The first one was obtained by retrieving the weighted graph. a community Neighborhood community ;

[0026] Calculate the first a community Community with neighbors The Middle a community Inter-module degree Assuming the merger of the first a community With the a community Calculate the new community after the merger self-modularity ;

[0027] Based on the new community self-modularity With the a community Community with neighbors The Middle a community Inter-module degree The comparison results determine whether the combined assumptions take effect or are overturned.

[0028] Repeat the iteration until it is no longer possible to merge communities, then stop. For each final community, calculate the mean of the encoded feature vectors of all users in each final community and organize them to construct a centroid feature matrix. .

[0029] Furthermore, based on the effective information entropy in information theory and the graph complexity in graph theory, the effective information entropy of a weighted graph is obtained as follows: The complexity of the weighted graph is Dynamic weight threshold It equals the effective information entropy divided by the complexity, i.e. Dynamic weight threshold The effectiveness and complexity of the weighted graph should be considered simultaneously.

[0030] Furthermore, the weight prediction network includes a modulation splicing layer, a multi-head cross-attention layer, and a nonlinear mapping layer;

[0031] The modulation and splicing layer will target the position Position encoding matrix is ​​generated through nonlinear mapping. and the centroid feature matrix Concatenate into a joint input matrix ;

[0032] Multi-head cross-attention layer includes Each attention branch will jointly input the matrix Enter the first The attention branch is generated through three different linear modulations. Query vectors , No. Key vectors and the Value Vector The Softmax function is used to convert the first... Query vectors With the Key vectors The dimension-normalized dot product result is transformed into the first dimension. Attention coefficient vector and the Value Vector Perform a dot product to generate the first... attention feature vectors ,Will The attention feature vectors output by each attention branch are weighted and summed using linear modulation to obtain the total attention feature vector. ,in, ;

[0033] The nonlinear mapping layer will focus on the total attention feature vector. Weight coefficient vectors are generated through nonlinear mapping. .

[0034] Compared with existing technologies, this invention collects channel parameter matrices from multiple users and generates a standard channel parameter matrix for each user through path completion and normalization. An adjacency matrix is ​​constructed based on the spatial distance between users. A graph autoencoder is used to perform feature aggregation and correlation capture on the standard channel parameter matrix, generating an encoded feature matrix and mapping it to a reconstructed adjacency matrix to verify whether an output is needed. A community-based partitioning clustering algorithm is used to iteratively merge the encoded feature vectors of each user, generating a final community set and calculating the centroid feature matrix. A weighted prediction network is used to expand the dimension mapping of the target location and concatenate it with the centroid feature matrix. Through a multi-head attention mechanism and the expanded dimension mapping, a weight coefficient vector is obtained, and the channel parameter matrix of each user is weighted and summed to obtain the channel parameter matrix of the target location. This accurately constructs a channel knowledge map. When a new user initiates a communication request, based on the new user's location and the channel knowledge map, a communication process between the base station and the new user is quickly established without the need for channel estimation. Attached Figure Description

[0035] Figure 1 Flowchart for building a wireless channel knowledge map;

[0036] Figure 2 Figure for a graph autoencoder model;

[0037] Figure 3 This is a schematic diagram of an information transmission mechanism;

[0038] Figure 4 This is a flowchart of a community-based partitioning clustering algorithm. Detailed Implementation

[0039] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0040] like Figure 1 As shown, a specific embodiment of the present invention discloses a fast communication method based on the construction of a wireless channel knowledge map, comprising the following steps:

[0041] The base station uses channel estimation algorithms from Measure the channel parameter matrix in the existing communication process of users at different locations to construct the user location set. and parameter matrix set ,in, and The first Location and channel parameter matrix of each user Total number of users;

[0042] Base station to the first Channel parameter matrix for each user Perform path completion and normalization to generate the first... Standard channel parameter matrix for each user ,use An adjacency matrix is ​​constructed based on the spatial distance relationships between individual users. ;

[0043] The base station uses a graph autoencoder, which transmits data through the encoder. The standard channel parameter matrix of each user is used for feature aggregation and nonlocal correlation capture to generate a coded feature matrix. Based on the feedback decoder, the encoded feature matrix is... Reconstructing the adjacency matrix through nonlinear mapping. Based on reconstruction loss Decision output encoding feature matrix Alternatively, update the encoder parameters and regenerate the encoded feature matrix. ;

[0044] The base station employs a community-based partitioning clustering algorithm, based on the encoded feature matrix. middle A weighted graph is constructed from the encoded feature vectors of each user, and each user is treated as a community. The similarity between communities is evaluated based on modularity, and the communities are iteratively merged until no further merging is possible, resulting in the final community set. Then, the centroid feature vector of each final community is calculated to generate the centroid feature matrix. ,in, For the first An endgame community. The final total number of communities;

[0045] The base station uses a weighted prediction network to determine the location of each target in the channel knowledge map. After performing dimension expansion mapping and the centroid feature matrix The feature matrix is ​​learned by splicing and multi-head attention mechanism. The contribution of eigenvectors with different centroids in the region is obtained again through dimension expansion mapping. The channel parameter matrix of each user with respect to the target location Channel parameter matrix The weighting coefficients will The channel parameter matrices of each user are summed using corresponding weighting coefficients to obtain the target location. Channel parameter matrix Construct a channel knowledge map;

[0046] The base station receives communication requests from new users and obtains the new users' locations. The nearest neighbor algorithm is used to match the location of new users with the channel knowledge map. The channel parameter matrix of the nearest location is used as the channel parameter matrix for the new user. It enables rapid establishment of communication processes with new users without the need for channel estimation.

[0047] Furthermore, the channel estimation algorithm is a prior art technique that inserts known pilot symbols when transmitting data and inversely derives the first channel signal based on the received pilot symbols when receiving data. Channel matrix between users and base stations The channel parameters in the model are based on the downlink channel model. Channel matrix between users and base stations Specifically as follows:

[0048] ,

[0049] in, and Base station and the Number of antennas per user For the first Number of paths per user For the first The user The gain of the path, and the first path The user Gain of each path Follows a mean of 0 and a variance of The cyclically symmetric complex Gaussian distribution, Represents the imaginary unit. For wireless communication frequencies, For the first The user The propagation delay of each path, , , , The first The user The arrival pitch angle, arrival azimuth angle, departure pitch angle, and departure azimuth angle of the path. and They represent the first The user The array response vectors of the path and the base station are calculated directly from the antenna array arrangement, which is a conventional calculation method in wireless communication and will not be described in detail in this application. , Since it is the conjugate transpose of a vector or matrix, therefore, the first... Channel parameter matrix for each user Specifically as follows:

[0050] ,

[0051] Based on the Channel parameter matrix for each user Based on the downlink channel model, the first... Channel matrix between users and base stations .

[0052] Furthermore, Users in different locations have different wireless communication environments, resulting in a different number of paths in each user's channel parameter matrix. Each user performs path completion separately to determine... Maximum number of paths among users If the number of paths is less than the maximum number of paths Below the user's channel parameter matrix, a 1×6 vector of all zeros is concatenated row by row, such that the number of rows is equal to the maximum number of paths. Maintain consistency and unity The dimension of the channel parameter matrix for each user is: ×6, to eliminate the dimensional differences in arrival elevation angle, arrival azimuth angle, departure elevation angle, departure azimuth angle, gain, and propagation delay in the channel parameter matrix, calculate respectively. The mean and standard deviation of the arrival pitch angle, arrival azimuth angle, departure pitch angle, departure azimuth angle, gain, and propagation delay for each user are calculated and Z-score normalized to ensure that the arrival pitch angle, arrival azimuth angle, departure pitch angle, departure azimuth angle, gain, and propagation delay for each user and each path are all within the range of 0 to 1. The standard channel parameter matrix for each user is used. Based on the concept of graph neural networks, each user is considered as a node, and the standard channel parameter matrix for each user is considered as a feature vector describing the node. The reciprocal of the spatial distance between two users is calculated and used as an edge connecting the two nodes to construct a node graph. Matrix form storage Edges between users are used to construct an adjacency matrix. Adjacency matrix The Line 1 List For the first The user and the first The reciprocal of the spatial distance between users.

[0053] like Figure 2 As shown, the graph autoencoder further includes an encoder and a feedback decoder;

[0054] The encoder uses graph convolution based on the adjacency matrix. For the Standard channel parameter matrix for each user Perform feature aggregation to obtain the first Embedded vectors of individual users By capturing the nonlocal correlation of embedding vectors from different users using Gaussian radial basis functions, an encoded feature matrix is ​​generated. ;

[0055] The feedback decoder uses a multilayer perceptron to encode the feature matrix. Dimensional expansion via linear modulation Furthermore, the ReLU function is employed to enhance the nonlinear representation and generate a reconstructed adjacency matrix. Based on reconstructing the adjacency matrix Adjacency Matrix Reconstruction loss Verification of the encoded feature matrix Does it adhere to the constraints of the correlation between the user's channel parameter matrix, and if so, the reconstruction loss? If the loss is less than or equal to the loss threshold, the encoded feature matrix is ​​output directly. If reconstruction loss If the loss exceeds the threshold, the reconstruction loss is calculated based on the gradient descent approach. Regarding the gradient of the encoder parameters, the encoder parameters are updated based on the negative gradient direction, and the encoded feature matrix is ​​regenerated. Among them, reconstruction loss Equivalent to reconstructing the adjacency matrix Adjacency Matrix Mean square error Encoded feature matrix L2 norm value With regularization factor The product of.

[0056] Furthermore, the encoder generates an encoded feature matrix. This includes the following steps:

[0057] Due to the adjacency matrix This reflects the correlation of channel parameter matrices between different users, hence the diagonal elements are 0. Then, graph convolution is used on the... Standard channel parameter matrix for each user Feature aggregation results in the loss of its own information, so the adjacency matrix is... Set all diagonal elements to 1 to generate a self-connected adjacency matrix. ;

[0058] like Figure 3 As shown, the information transfer mechanism of graph convolution is used to process the first... Standard channel parameter matrix for each user conduct Secondary feature aggregation generates the th... Embedded vectors of individual users , among which, the The input for the second feature aggregation is the first... The user The aggregation channel parameter matrix of the second order. The output is the first... The user The aggregation channel parameter matrix of the second order. The details are as follows:

[0059] ,

[0060] in, For activation function, For the first Learnable weight matrix during secondary feature aggregation For self-connected adjacency matrix The degree matrix is ​​used to connect the self-connected adjacency matrix. Normalization is performed to ensure fairness and stability during information transmission. For the first The degree vector of each user , For self-connected adjacency matrix The first in Line 1 If This reflects the first The user and the first The reciprocal of the spatial distance between users, i.e. ,like ,but , The 0th aggregation channel parameter matrix That is, the first Standard channel parameter matrix for each user , No. The aggregation channel parameter matrix of the second order. That is, the first Embedded vectors of individual users Since graph convolution utilizes a message-passing mechanism to aggregate features among users with edges, the total number of feature aggregations increases. As the number of users increases, their characteristics gradually become similar, causing an oversmoothing problem. This embodiment sets... To avoid oversmoothing issues;

[0061] The Gaussian radial basis function is used to calculate the first... The user and the first Encoded response between users The details are as follows:

[0062] ,

[0063] Among them, the encoded response Reflecting the Embedded vectors of individual users With the Embedded vectors of individual users Feature associations between them The scaling parameters are pre-set for the Gaussian radial basis functions. This represents an exponential function with the natural constant as its base, which encodes the response. Corresponding to the number Line 1 Columns, generate encoded feature matrix Since graph convolution is based on the adjacency matrix Feature aggregation is performed because information transmission between users with low similarity in channel parameter matrices is weak. Gaussian radial basis functions can effectively capture nonlocal feature correlations.

[0064] like Figure 4 As shown, the community-based partitioning clustering algorithm further includes the following steps:

[0065] Obtain the encoded feature matrix The Middle Encoded feature vector of each user With the Encoded feature vector of each user , among which, the Encoded feature vector of each user Encoding feature matrix The row or number The transpose of the column will... The user and the Each user is considered a vertex, and the cosine similarity of the encoded feature vectors is used to construct the first... The user and the first Weight edge between users The details are as follows:

[0066] ,

[0067] in, Represents the encoded feature vector transpose;

[0068] Considering the total number of users in different application scenarios It is changing, based on the total number of users. Set dynamic weight threshold ,right Pruning is performed on weighted edges between users, removing non-critical weighted edges to reduce computational complexity; that is, retaining edges with weights greater than the dynamic weight threshold. Remove the weighted edges and delete the rest that are less than or equal to the dynamic weight threshold. Construct a weighted graph using the weighted edges;

[0069] Treating each user in the weighted graph as an independent community, and defining communities with weighted edges as neighboring communities, for the ... a community The first one was obtained by retrieving the weighted graph. a community Neighborhood community ;

[0070] Calculate the first a community Community with neighbors The Middle a community Inter-module degree The details are as follows:

[0071] ,

[0072] in, For the first a community With the a community The correlation weight between them , and These are the weighted edges of the weighted graph, and the first edge is the edge weight. a community The sum of association weights with all neighboring communities, and the first a community The sum of association weights with all neighboring communities, where the first... a community With the a community Association weight equal to the a community Includes all users and the first The sum of weighted edges among all users in a community;

[0073] Assuming the merger of the first a community With the a community Get New Community , computing new community self-modularity The details are as follows:

[0074] ,

[0075] in, For the new community The weighted sum of all neighboring communities, of which the new community The neighboring community is the first a community Neighborhood community With the a community Neighborhood community The union of neighboring communities Community with neighbors If there are neighboring communities in the same area, then the new community The association weight with the same neighboring community is equal to the first a community The association weight with the same neighboring community plus the first a community Association weight with the same neighboring communities;

[0076] Ruoxin Community self-modularity Greater than the a community Community with neighbors The Middle a community Inter-module degree If the merger takes effect, a new community will be formed. If the new community self-modularity Less than or equal to the a community Community with neighbors The Middle a community Inter-module degree Then the assumption is overturned, and the original assumption is maintained. a community With the a community ;

[0077] Repeat the iteration until all communities can no longer be merged, then stop. A final community, building a collection Calculate the mean of the encoded feature vectors of all users in each final community and use it as the centroid feature vector. Concatenate the centroid feature matrix row by row. ,in, For the first The centroid feature vector of each end-game community.

[0078] Furthermore, dynamic weight threshold Based on the effective information entropy in information theory and the graph complexity design in graph theory, for existing A weighted graph with weighted edges has an effective information entropy of . The complexity of the weighted graph is Therefore, dynamic weight threshold It equals the effective information entropy divided by the complexity, i.e. Dynamic weight threshold The effectiveness and complexity of the weighted graph should be considered simultaneously.

[0079] Furthermore, the weight prediction network includes a modulation splicing layer, a multi-head cross-attention layer, and a nonlinear mapping layer;

[0080] The modulation and splicing layer will target the position Dimensionality expansion is achieved through linear modulation, and the dimension and centroid feature matrices are generated by ReLU function mapping. Same position encoding matrix splicing the centroid feature matrix With position encoding matrix Generate joint input matrix ;

[0081] Multi-head cross-attention layer includes Several parallel attention branches will jointly input the matrix Enter the first Each attention branch modulates the joint input matrix through three different linear modulations. Adjusted to the number Query vectors , No. Key vectors and the Value Vector The Softmax function is used to convert the first... Query vectors With the Key vectors The dimension-normalized dot product result is transformed into the first dimension. Attention coefficient vector , will the Attention coefficient vector With the Value Vector Perform a dot product to generate the first... attention feature vectors The details are as follows:

[0082] ,

[0083] in, For the first Key vectors Dimensions ,Will The attention feature vectors output by each attention branch are weighted and summed using linear modulation to obtain the total attention feature vector. ;

[0084] The nonlinear mapping layer will focus on the total attention feature vector. The dimension is adjusted to 1× through linear modulation. And then generate the weight coefficient vector again through the ReLU function mapping. ,in, For the first The weighting coefficient of each user.

[0085] This invention discloses a fast communication method based on the construction of a wireless channel knowledge map. It collects channel parameter matrices from multiple users and generates a standard channel parameter matrix for each user through path completion and normalization. An adjacency matrix is ​​constructed based on the spatial distance between users. A graph autoencoder is used to perform feature aggregation and correlation capture on the standard channel parameter matrix, generating an encoded feature matrix and mapping it to a reconstructed adjacency matrix to verify output. A community-based partitioning clustering algorithm iteratively merges the encoded feature vectors of each user to generate a final community set and calculates the centroid feature matrix. A weighted prediction network is used to expand the dimension of the target location and concatenate it with the centroid feature matrix. A multi-head attention mechanism and the expanded dimension mapping are used to obtain a weight coefficient vector, which is then weighted and summed to obtain the channel parameter matrix of each user, resulting in the channel parameter matrix of the target location. This achieves accurate construction of the channel knowledge map. Based on the channel knowledge map, when a new user communicates with a base station, channel estimation is not required; the channel parameter matrix of the new user is directly matched based on the new user's location, enabling fast communication between the new user and the base station.

[0086] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A fast communication method based on wireless channel knowledge map construction, characterized in that, Includes the following steps: Collected from the communication process The location and channel parameter matrices of each user are used. Path completion and normalization are performed on the channel parameter matrix of each user to obtain a standard channel parameter matrix. Then, using... An adjacency matrix is ​​constructed using the spatial distances between users. Total number of users; Using a graph autoencoder, for The standard channel parameter matrix of each user is used for feature aggregation and correlation capture to generate a coding feature matrix. The coding feature matrix is ​​then mapped to a reconstructed adjacency matrix, and a decision on whether to output the coding feature matrix is ​​made based on the reconstruction loss. A community-based partitioning clustering algorithm is adopted. A weighted graph is constructed based on the encoded feature matrix and each user is regarded as a community. The similarity between communities is evaluated based on the modularity and iteratively merged until it can no longer be merged, thus obtaining the final community set. The centroid feature vector of each final community is calculated to generate the centroid feature matrix. A weighted prediction network is used to expand the dimension of the target location and then concatenate it with the centroid feature matrix. A multi-head attention mechanism is then used to obtain the weight coefficient vector through the expanded dimension mapping. The channel parameter matrices of each user are weighted and summed to obtain the channel parameter matrix of the target location. The channel parameter matrices of all target locations in the channel knowledge map are then calculated to construct the channel knowledge map. Receive communication requests and new user locations from new users. The nearest neighbor algorithm is used to match the channel parameter matrix of new users from the channel knowledge map. To establish a communication process with new users.

2. The fast communication method based on wireless channel knowledge map construction as described in claim 1, characterized in that, Community-based partitioning clustering algorithms include the following steps: The first The user and the first Each user is considered a vertex, and the _th user in the encoded feature matrix is ​​used as the vertex. The encoded feature vector of the user and the first The cosine similarity of the encoded feature vectors of each user is used to construct the first... The user and the first Weighted edges between users ; Based on total number of users Set dynamic weight thresholds and... The weighted edges between users are pruned, retaining weighted edges greater than the dynamic weight threshold, and a weighted graph is constructed. Treating each user in the weighted graph as an independent community, for the... The community was retrieved based on the weighted graph. Each community has neighboring communities with weighted edges. Construct the first... Neighborhood community groups of a community; Calculate the first The community and neighboring communities are concentrated in the first The inter-module degree of each community, assuming the merging of the first... The community and the first For each community, calculate the self-modularity of the merged new community; Based on the self-modularity and the first The community and the first The comparison results of the intermodality of each community determine whether the assumptions for merging or overturning are valid. Repeat the iteration until it is no longer possible to merge communities, then stop. For each final community, calculate the mean of the encoded feature vectors of all users in each final community and organize them to construct a centroid feature matrix.

3. The fast communication method based on wireless channel knowledge map construction as described in claim 2, characterized in that, Based on the effective information entropy in information theory and the graph complexity in graph theory, the effective information entropy of a weighted graph is obtained as follows: The complexity of the weighted graph is The dynamic weight threshold is equal to the effective information entropy divided by the complexity.

4. The fast communication method based on wireless channel knowledge map construction as described in claim 1, characterized in that, The image encoder includes an encoder; The encoder uses graph convolution to perform graph convolution on the first-order graph based on the adjacency matrix. Feature aggregation is performed on the standard channel parameter matrix of the nth user to obtain the nth user. The embedding vectors of each user are used to capture the nonlocal correlation of the embedding vectors of different users through Gaussian radial basis functions, thereby generating an encoded feature matrix.

5. The fast communication method based on wireless channel knowledge map construction as described in claim 4, characterized in that, The encoder sets all diagonal elements of the adjacency matrix to 1 to generate a self-connected adjacency matrix, and calculates the degree matrix of the self-connected adjacency matrix; it then uses the information transfer mechanism of graph convolution to process the first... Standard channel parameter matrix for each user Secondary feature aggregation generates the th... The embedding vector of the nth user; the Gaussian radial basis function is used to calculate the nth... The user and the first The encoded responses between users are arranged in order of... Line 1 Columns, generate encoded feature matrix, .

6. The fast communication method based on wireless channel knowledge map construction as described in claim 1, characterized in that, The graph autoencoder also includes a feedback decoder; The feedback decoder uses a multilayer perceptron to map the encoded feature matrix to a reconstructed adjacency matrix. It calculates the reconstruction loss between the reconstructed adjacency matrix and the adjacency matrix and compares it with a loss threshold. If the loss loss is less than or equal to the loss threshold, the encoded feature matrix is ​​output. If the loss loss is greater than the loss threshold, the encoder parameters are updated based on gradient descent and the encoded feature matrix is ​​regenerated. The reconstruction loss is equal to the mean square error between the reconstructed adjacency matrix and the adjacency matrix plus a regularization term calculated based on the encoded feature matrix.

7. The fast communication method based on wireless channel knowledge map construction as described in claim 1, characterized in that, right Each user performs path completion separately to determine... Maximum number of paths per user Each path has a number of paths less than the maximum number of paths. The user's channel parameter matrix is ​​padded with a vector of all zeros so that the number of rows is equal to the maximum number of paths. Consistent, calculate separately The mean and standard deviation of the arrival pitch angle, arrival azimuth angle, departure pitch angle, departure azimuth angle, gain, and propagation delay for each user are calculated and normalized using Z-Score to obtain... The standard channel parameter matrix for each user.

8. The fast communication method based on wireless channel knowledge map construction as described in claim 1, characterized in that, Treat each user as a node, calculate the reciprocal of the spatial distance between two users using their positions, and use this as an edge connecting the two nodes. Construct a node graph and store it in matrix form. Construct an adjacency matrix by finding the edges between n users. The nth edge of the adjacency matrix is... Line 1 Listed as number The user and the first The reciprocal of the spatial distance between users.

9. The fast communication method based on wireless channel knowledge map construction as described in claim 1, characterized in that, Based on the downlink channel model, the first The channel parameter matrix for each user includes The arrival pitch angle, arrival azimuth angle, departure pitch angle, departure azimuth angle, gain, and propagation delay of the path. For the first The number of paths for each user, the first The channel parameters for each user are obtained through a channel estimation algorithm.

10. The fast communication method based on wireless channel knowledge map construction as described in claim 1, characterized in that, The weighted prediction network generates a position encoding matrix for the target location through a nonlinear mapping and concatenates it with the centroid feature matrix to form a joint input matrix; the multi-head cross-attention layer includes The attention branch inputs the joint input matrix into the first attention branch. The attention branch generates the first attention branch. A weighted sum of attention feature vectors. Each attention feature vector is used to generate a weight coefficient vector again through a nonlinear mapping, where... .

Images