Rapid construction method and system for wireless channel knowledge map

Abstract

The present invention discloses a rapid construction method and system for a wireless channel knowledge map. The present invention uses image restoration technology to construct a channel knowledge map, and, by means of recovering a corresponding pixel value in an image matrix, predicts channel knowledge of a specific position, so that channel related knowledge of a user equipment at any position in a target area can be acquired. By means of combining the similarity of an image restoration problem and a channel knowledge map construction problem, and using a Laplacian pyramid frequency band decomposition framework, an input environment map can be decomposed at different frequencies. According to the characteristics of different frequency components, a corresponding sub-network is designed for feature extraction, and finally, the channel knowledge map is reconstructed by means of inverse operation of the Laplacian pyramid. The wireless channel knowledge map construction method provided by the present invention not only implements higher construction precision at lower computational complexity, but also has strong generalization ability for various wireless communication scenarios.

Description

A method and system for rapidly constructing a wireless channel knowledge map Technical Field

[0001] The present invention belongs to the field of communication technology and relates to a channel knowledge map construction method and system based on an image restoration method. Background Art

[0002] With the increasing number and density of connected devices, the expansion of antenna array sizes, and the wider use of bandwidth, 6G will involve extremely large-dimensional wireless channels. Traditional pilot-based channel training and feedback methods for acquiring real-time channel state information (CSI) incur excessive overhead. It is worth noting that the wireless propagation environment, such as the geometric location relationships in a city or terrain map, is not only static but also a key factor affecting channel parameters and the performance of wireless communication systems. Therefore, context-aware wireless communications have attracted significant research interest and attention in academia and industry. Channel knowledge maps (CKMs) play a key role in context-aware wireless communications. They serve as a specialized database that identifies the precise locations of transmitters and receivers. This database stores essential channel-related details, providing valuable information such as channel gain, shadowing, arrival / departure angles, and channel impulse response. Effectively leveraging the propagation environment to acquire this channel-related information can address the high complexity of real-time CSI acquisition. A typical example of a CKM is a channel gain map (CGM), which is used to predict the channel gain at specific locations in a target area.

[0003] Traditional approaches rely on data-driven interpolation methods and model-driven parameter fitting algorithms. However, existing CKM construction methods based on interpolation and model parameter fitting do not fully consider the specific wireless propagation environment of the actual channel. As a result, implementation limitations exist, such as the large number of measurement points required, large storage capacity, high computational complexity, and low estimation accuracy.

[0004] Summary of the Invention

[0005] Purpose of the invention: In response to the shortcomings of the existing technology, the purpose of the present invention is to provide a method and system for quickly constructing a wireless channel knowledge map, which can obtain channel knowledge information corresponding to user equipment at specific locations within the base station target area, while ensuring reconstruction accuracy, and can further reduce time complexity and storage complexity compared to existing methods.

[0006] Technical Solution: To achieve the above-mentioned purpose, the present invention proposes a method and system for rapidly constructing a wireless channel knowledge map. Through an end-to-end Laplacian Pyramid (LP)-based CKM construction solution, the CKM construction problem is transformed into an image-to-image restoration task. Corresponding subnetworks are designed based on the importance of different components, and self-attention mechanisms and cross-attention mechanisms are further introduced to encode global structural information to improve the reconstruction accuracy and generalization ability of the model. The method for rapidly constructing a wireless channel knowledge map includes the following steps:

[0007] Obtain the environmental information of the target area and convert it into a 2D image format through spatial discretization and pixel-level grayscale conversion, thereby converting the channel knowledge map construction problem into an image restoration task;

[0008] Use the Laplacian pyramid to decompose the environment map into different frequency bands to obtain different frequency components;

[0009] Constructing a channel knowledge map reconstruction network model; the network model designs different sub-networks for feature extraction for different frequency components of the environment map, and after the different frequency components are output by their respective sub-networks, the channel knowledge map is reconstructed using the inverse operation of the Laplacian pyramid;

[0010] Given a training set containing an environment map and its corresponding channel knowledge map, the constructed network model is trained using an end-to-end supervised training method to obtain the optimal parameters of the model;

[0011] The trained model is used to predict the channel knowledge information of user devices at specific locations in the target area.

[0012] Preferably, the transformation of the channel knowledge map construction problem to the image restoration problem is to treat each specific position as a pixel point of the image matrix, and the channel knowledge corresponding to the specific position as the pixel value of the corresponding pixel point, and transform the channel knowledge estimation problem of each specific position into a pixel restoration problem between images.

[0013] Preferably, the spatial discretization method grids the environmental information in the spatial dimension and is implemented according to a minimum distance criterion.

[0014] Preferably, the grayscale conversion is implemented at the pixel level, and the channel knowledge value is converted into an interval range of 0 to 1 using minimum-maximum normalization.

[0015] Preferably, the frequency band decomposition utilizes a Gaussian pyramid to obtain a Laplacian pyramid, thereby obtaining different frequency components. For a pair of an environment map and a channel knowledge map, the difference in low-frequency components is more pronounced than in high-frequency components; the sub-network corresponding to the low-frequency components has more feature extraction levels than the sub-network corresponding to the high-frequency components.

[0016] Preferably, the above-mentioned model mainly includes:

[0017] Reconstruction of low-frequency components in the Lth layer: The input environment map is decomposed into sub-images of different frequency bands, and their specific properties are used to complete feature extraction and reconstruction. L First, it undergoes a point convolution to complete the dimensional expansion in the depth direction. Then, it is input into multiple stacked lightweight residual dilated convolutions (LRDC) modules to deepen the feature extraction network and obtain multi-scale feature fusion. Next, it is input into the multi-head self-attention (MHSA) module to encode the global random structure information into the local features of the spatial dimension and learn a richer hierarchical feature representation. Finally, the number of color channels in the feature map is reduced to the original size to obtain the output result.

[0018] Refinement of high-frequency components in layer L-1: The low-frequency components and their reconstruction results are fused to gradually guide the refinement of high-frequency components to obtain components of different frequencies for reconstructing the channel knowledge map. First, the low-frequency environment sub-map I L and predicted channel knowledge subgraph Then, the low frequency component I L , the reconstruction results of low-frequency components and the high-frequency component r of the L-1 layer L-1 The extracted multi-scale feature maps are then input into the Multi-Head Cross-Covariance Attention (MHCCA) module. Finally, the high-frequency component r of the L-1 layer is obtained. L-1 The refined results and the attention map used to refine the high-frequency components of the L-2 layer. For the remaining frequency components R=[r0,...,r L-2], through iterative upsampling, concatenation, LRDC modules, and MHSA or MHCCA modules, all high-frequency components of the Laplacian pyramid are gradually refined, and prediction results for different components are obtained. The channel knowledge map is reconstructed through the inverse operation of the Laplacian pyramid. The number of LRDC modules used in the low-frequency and high-frequency sub-networks decreases, and each sub-network can use MHSA or MHCCA modules without considering computational complexity.

[0019] Preferably, the LRDC module consists of a depthwise separable convolutional layer, an instance normalization layer, a pointwise convolutional layer, and a GELU activation function layer for feature extraction and representation. The MHSA module consists of a normalization layer, a multi-head self-attention layer, a pointwise convolutional layer, and a GELU activation function layer. The MHCCA module consists of a normalization layer, a multi-head cross-covariance attention layer, a pointwise convolutional layer, and a GELU activation function layer for encoding global context information.

[0020] As a preference, the Adam optimization algorithm and end-to-end supervised training are used to train the parameters of the above model to minimize the cost function. The optimal model parameters mainly include the weights and biases of the point convolution layer, the depthwise separable dilated convolution layer, the multi-head self-attention layer, and the multi-head cross-covariance attention layer. The cost function is expressed as: Among them, K is the total number of training set samples, is the reconstructed channel knowledge map, It is a real channel knowledge map. Represents the constructed network model, Θ is The set of parameters to be learned in , represents the environment map, e represents the propagation environment, x n represents the user device location, and ||·||2 represents the Euclidean norm.

[0021] The present invention also provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program implements the method for rapidly constructing a wireless channel knowledge map when loaded into the processor.

[0022] Based on the same inventive concept, the present invention also provides a wireless communication system, including a base station and multiple user terminals, wherein the base station or user terminal is provided with: an information processing module, used to obtain environmental information of the target area, and convert the environmental information into a 2D image format through a spatial discretization method and pixel-level grayscale conversion, thereby converting the channel knowledge map construction problem into an image restoration task; a frequency band decomposition module, used to use a Laplace pyramid to decompose the environmental map in different frequency bands to obtain different frequency components; a network model construction module, used to construct a channel knowledge map reconstruction network model; the network model designs different subnetworks for different frequency components of the environmental map for feature extraction, and after the different frequency components are output by their respective subnetworks, the channel knowledge map is reconstructed with the help of the inverse operation of the Laplace pyramid; a model training module, used to train the constructed network model using an end-to-end supervised training method given a training set containing an environmental map and its corresponding channel knowledge map to obtain the optimal parameters of the model; and a prediction module, used to use the trained model to predict the channel knowledge information of the user equipment at a specific location in the target area.

[0023] Beneficial effects: Compared with the existing technology, the present invention transforms the channel knowledge map construction problem into an image restoration problem. Based on the reversible frequency band decomposition architecture of the Laplace pyramid, corresponding sub-networks are designed for frequency components with different characteristics, and the reconstruction of low-frequency components and the refinement of high-frequency components are completed. Finally, the reconstruction of the channel knowledge map is completed through the inverse operation of the Laplace pyramid. The proposed method for quickly constructing a wireless channel knowledge map can obtain the channel knowledge information of user devices at potential locations in the target area with low time complexity and storage complexity while ensuring the reconstruction accuracy. The obtained channel knowledge can further assist in the acquisition of real-time channel state information of large-scale wireless communications, improve system performance, and thus further improve the overall transmission efficiency of the system. In addition, the method proposed in the present invention has strong generalization ability and is applicable to different large-scale wireless communication scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] FIG1 is a flow chart of a rapid construction of a channel knowledge map according to an embodiment of the present invention;

[0025] FIG2 is a framework diagram of a channel knowledge map rapid construction model according to an embodiment of the present invention;

[0026] FIG3 is a schematic diagram of an LRDC module in an embodiment of the present invention;

[0027] FIG4 is a schematic diagram of the MHSA module and the MHCCA module in an embodiment of the present invention;

[0028] FIG5 is a schematic diagram showing the effect of the number of attention heads on the performance of the embodiment when the number of Laplacian pyramid layers is 3;

[0029] FIG6 is a schematic diagram showing a comparison of the channel gain map reconstruction performance between the construction method according to an embodiment of the present invention and the existing method. DETAILED DESCRIPTION

[0030] The technical solutions provided by the present invention will be described in detail below with reference to specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention.

[0031] As shown in Figures 1 and 2, an embodiment of the present invention discloses a method for quickly constructing a wireless channel knowledge map, which converts the channel knowledge map construction problem into an image-to-image restoration task, and uses the Laplacian pyramid to extract frequency components of different importance, and then designs corresponding sub-networks according to the importance of different components to improve the overall reconstruction speed, and introduces self-attention mechanism and cross-attention mechanism to encode global structural information, thereby improving the reconstruction accuracy and generalization ability of the model. The specific steps include: (1) obtaining the environmental information of the target area, and converting the environmental information into a 2D image format through a spatial discretization method and pixel-level grayscale conversion, thereby converting the channel knowledge map construction problem into an image restoration task. (2) Using the Laplacian pyramid to decompose the environmental map into different frequency bands, obtain different frequency components for feature extraction and reconstruction. (3) Reconstructing the low-frequency components, the input environmental map is decomposed into sub-graphs of different frequency bands, and their specific attributes are used to complete feature extraction and reconstruction. Low-frequency sub-graph I L First, it undergoes a 1×1 convolution to complete the dimensionality expansion in the depth direction. Then, it is input to the LRDC module to deepen the feature extraction network and obtain the fusion of multi-scale features. Then, it is input to the MHSA module to encode the global random structure information into the local features of the spatial dimension and learn a richer hierarchical feature representation. Then, the number of color channels in the feature map is reduced to c to obtain the result (4) Refine the high-frequency components and convert the low-frequency sub-image I L and reconstructed The fusion result gradually guides the refinement of high-frequency components R = [r0, r1, r2, ..., r L-1 ] to obtain the components of different frequencies for reconstructing CGM. First, the low-frequency geometric position subgraph and predicted channel knowledge subgraph Bilinear upsampling is performed respectively, and the spatial resolution is adjusted to and the high frequency component r of the (L-1)th layer L-1 The resolution is the same as that ofL 、 and r L-1 Splice and define it as Then, it is input to the LRDC module, and its output is expressed as Then, the extracted multi-scale feature map is input into the MHCCA module, and the attention operation is performed along the feature dimension. Finally, the high-frequency component r is obtained. L-1 Through the above steps, all high-frequency components of the Laplacian pyramid can be gradually refined and the prediction results of different components can be obtained. exist and With the help of the inverse operation of the Laplace pyramid, the channel knowledge map can be reconstructed In this example, N1, N2, N3 and N4 are set to 4, 3, 3, and 2 respectively. (5) The cost function of the entire network model is designed to be the mean square error between the channel knowledge map output by the network and the actual channel knowledge map, which is in is the reconstructed channel knowledge map, It is a real channel knowledge map. Represents the constructed network model, Θ is The parameter set to be learned in , ||·||2 represents the Euclidean norm. (6) The data set is divided into a training set (40,000), a validation set (8,000), and a test set (8,000) in a ratio of 5:1:1. (7) The above model is trained using the Adam optimization algorithm and an end-to-end supervised training method. The batch size set during the training process is 15, and the learning rate is set to 0.0001. This training strategy is used to traverse the training set for 50 rounds. During the training process, the weight parameters of the model are adjusted using the validation set, and the final performance of the model is tested using the test set. (8) The above trained model is used to quickly reconstruct the channel knowledge map of the target area in the wireless communication system. According to the input environment map, different frequency components are obtained through the Laplacian pyramid, and then the channel knowledge map in the target area can be output after inputting the above model.

[0032] FIG3 illustrates the specific framework components of the LRDC module; FIG4 illustrates the specific framework components of the MHSA module and the MHCCA module.

[0033] In order to enable those skilled in the art to better understand the solution of the present invention, the specific steps of the embodiment of the present invention are described below with reference to a specific scenario, taking the prediction of channel gain as an example.

[0034] 1. System

[0035] Consider a square coverage area Consider a wireless communication scenario in Figure 1, which includes a base station (BS) and N user equipment (UE). The signal power attenuation observed at the UE may be caused by various factors, such as propagation loss in different paths, reflection and diffraction from buildings, waveguide effects in the street, and obstructions. Among these influences, the relatively slow-changing parts together constitute the channel gain function, which is defined as G L (e,x n ,f), describes the frequency In this case, UE location The large-scale signal attenuation measured at the location. It is worth noting that G L (e,x n ,f) is significantly affected by the propagation environment, denoted by e. In addition, small-scale effects are usually modeled as a complex Gaussian random variable H with unit variance. The baseband signal received at the UE can be described as:

[0036] Where X is a number with power P X The transmitted signal is Z, and Z is the additive noise with a unilateral power spectrum density of N0. The average received energy of each symbol is

[0037] Where B represents the bandwidth of the signal. The signal-to-noise ratio (SNR) at the input of the UE baseband processor is:

[0038] For ease of processing, channel gain is defined using the decibel (dB) scale:

[0039] This definition describes the received power at the user location x n Accurately estimate P at each location g It is the key to building a refined CGM.

[0040] 2. Problem Statement

[0041] For any x n ∈A, the goal of this invention is to predict the corresponding P defined in formula (4) g To do this, we need to construct a CGM, defined as f Θ , which provides the n To the corresponding P g The mapping is:

[0042] Constructing this mapping (5) is tricky, G L (e,x n,f) is affected by the actual propagation environment, while analytical path loss and ,shadowing models can only provide a rough approximation. Therefore, it is challenging to ,accurately obtain the global and complete CGM corresponding to the target ,area.

[0043] Stochastic modeling ignores environmental information, such as the morphology of streets, buildings, and spatial geometric relationships. In contrast, in image restoration tasks, the focus of feature extraction and representation is to learn the intrinsic structure of the image. The present invention applies image restoration technology to the CGM construction problem. Let the input of the neural network be the environment map, in this case the input geometric position map, and the output be the desired CGM. The difference between the two is the channel gain to be estimated, which can actually be regarded as the content that needs to be repaired in the image restoration task. The CGM construction problem is converted into an image restoration task, and the channel gain of the UE position is predicted by restoring the corresponding pixels in the image matrix. Then, the present invention proposes an efficient CGM construction and reconstruction network, which is defined as And express Θ as In supervised training mode, a set of training samples is provided, which includes the input environment map And the corresponding output CGM is defined as Where k = 1,...,K, K is the number of training samples. Environment map It mainly contains the environment e, such as base station locations, buildings, roads, vehicles, etc. This information is saved in PNG image format. The goal of model training is to continuously The parameter set Θ, using As prior knowledge, to effectively reconstruct the desired CGM, namely:

[0044] in, is the reconstructed CGM. Therefore, from the perspective of image restoration, the CGM construction problem can be expressed as:

[0045] where ||·||2 denotes the Euclidean norm. The parameter set Θ can be learned by a gradient descent optimizer, such as the Adaptive Momentum Estimation (Adam) optimizer.

[0046] 3. CGM reconstruction method based on Laplacian pyramid

[0047] 1. Spatial discretization and grayscale conversion of geometric position map and channel gain

[0048] For the target location area A, discretize along its geometric space dimension, that is, take Δ x and Δ y is the minimum interval unit and is meshed into a grid with N xRow and N y 2D vector graph of columns. Each spatial grid is represented by Γ i,j , where i = 1, 2, ..., N x ,j=1,2,...,N y , the (i,j)th spatial grid can be expressed as: Γ i,j :=[iΔ x ,jΔ y ] T (8)

[0049] For the convenience of expression, let represents the index set of all UE position coordinates contained in the (i, j)th spatial grid. This spatial discretization process is performed according to the minimum distance criterion, that is, if and only if Only then When the equation holds true, arbitrarily determine or

[0050] By discretizing the geometric position of the UE, the spatial position set The channel gain vector is rearranged into a 2D channel gain matrix It should be noted that based on the above grid allocation method, if N<N x N y , then the vector P g The component values ​​of (f) cannot completely fill the matrix P g (f). In this case, fill the matrix P with 0 g The blank part in (f). The channel gain of the (i, j)th spatial grid is defined as [P g (f)] i,j =P g (Γ i,j ,f). In addition, when considering multiple frequencies, the 2D CGM can be spliced ​​together in the frequency dimension to form a tensor representation Right now:

[0051] Obviously, if the discretized geometric position map has a high spatial resolution, that is, when Δ x and Δ y Small enough, then for all Both Established. The channel gain for continuous position coordinates can be well approximated as:

[0052] Note that in Equation (10), when Δx and Δ y When both are close to 0, each set A i,j There is at most one element. Therefore, the summation symbol on the left side of Equation (10) is only for symbolic convenience, and there is at most one term in the summation. The spatial discretization transformation allows the channel gain estimates at continuous positions to be converted into pixel or grid predictions, which also facilitates the design of subsequent neural networks.

[0053] Through the above operations, corresponding to the spatial grid Γ i,j The channel gain P g (Γ i,j ,f) is regarded as the value of the corresponding pixel in the image matrix and needs to be grayscale converted before further processing. To this end, the minimum-maximum normalization is used to convert its value to the range of 0 to 1. The channel gain of the grayscale level can be expressed as:

[0054] 2. Laplace pyramid band decomposition

[0055] Given an input image I0 of size h×w, we first obtain a low-pass estimate through the Laplacian pyramid Each pixel is a weighted average of neighboring pixels based on an octave Gaussian filter. In order to ensure the reversibility of image reconstruction, the Laplacian pyramid retains high-frequency residual information. in represents the result of upsampling I1. To further reduce the input resolution, the Laplacian pyramid iteratively performs the above operations on I1, generating a series of low-frequency and high-frequency components. The reversible and closed-form decomposition properties of the Laplacian pyramid address the irreversibility of upsampling and downsampling. Furthermore, the Laplacian pyramid can decompose the high-resolution input into multiple sub-images with different spatial resolutions, which brings the benefit of fusing multi-scale feature information.

[0056] 3. CGM reconstruction network based on Laplacian pyramid

[0057] This paper proposes an end-to-end Laplacian Pyramid-based CGM Reconstructed Network (LPCGMN) to reduce computational complexity while maintaining competitive performance. The proposed Laplacian Pyramid-based CGM reconstruction network consists of the following two parts:

[0058] 1) Reconstruction of low-frequency components: The input environment map is decomposed into sub-images of different frequency bands, and their specific properties are used to complete feature extraction and reconstruction. LFirst, it undergoes a 1×1 convolution to complete the dimensionality expansion in the depth direction. Then, it is input to the LRDC module to deepen the feature extraction network and obtain the fusion of multi-scale features. Then, it is input to the MHSA module to encode the global random structure information into the local features of the spatial dimension and learn a richer hierarchical feature representation. Then, the number of color channels in the feature map is reduced to c to obtain the result

[0059] 2) Refined processing of high frequency components: I L and The fusion result gradually guides the refinement of high-frequency components R = [r0, r1, r2, ..., r L-1 ] to obtain the components of different frequencies for reconstructing CGM. First, the low-frequency environment sub-image and the predicted channel gain subgraph Bilinear upsampling is performed respectively, and the spatial resolution is adjusted to and the high frequency component r of the (L-1)th layer L-1 The resolution is the same as that of L 、 and r L-1 Splice and define it as Then, it is input to the LRDC module, and its output is expressed as Then, the extracted multi-scale feature map is input into the MHCCA module, and the attention operation is performed along the feature dimension. Finally, the high-frequency component r is obtained. L-1 Therefore, through the above steps, all high-frequency components of the Laplacian pyramid can be gradually refined and the prediction results of different components can be obtained. exist and With the help of the inverse operation of the Laplacian pyramid, the channel gain map can be reconstructed

[0060] 4. Three modules in the CGM reconstruction network based on the Laplacian pyramid

[0061] 1) LRDC module: It mainly uses depthwise separable convolution and dilated convolution to obtain multi-scale features with fewer model parameters. Specifically, assuming that d is defined as the dilation rate and k is the size of the ordinary convolution kernel, the actual size of the dilated convolution kernel is k′=k+(k-1)(d-1), and the receptive field of the (i+1)th layer can be expressed as: RF i+1 =RF i +(k′-1)×S i (13)

[0062] Among them, stride i The sampling interval of the i-th convolution layer, S i is the product of all strides in the first (I-1) layer, RF i It is defined as the receptive field of the i-th layer. When calculating the receptive field of the (i+1)-th layer, RF1 is initialized to 1 by default. Assuming x[n] represents the input, the output can be written as:

[0063] Where w[k] is a filter of length K. Consider an input feature X of dimension H×W×C, the LRDC module outputs It can be expressed as:

[0064] Among them, IN(·) is the instance normalization layer, GELU(·) is the GELU activation function, and DWSDConv d (·) represents a 3×3 depthwise separable convolution with a dilation rate of d, and PWConv(·) is a pointwise convolution.

[0065] 2) MHSA module: given input d m Defined as the number of image patches, d n Defined as the feature dimension of each image block, for z j Apply three different linear transformations: k j =z j W k ,j=1,...,d n (16) q j =z j W q ,j=1,...,d n (17) v j =z j W v ,j=1,...,d n (18)

[0066] in, and are the key, query, and value vectors respectively. and Represent the trainable transformation matrix respectively. For the sake of clarity, formulas (16)-(18) are expressed in matrix form: K = ZW k (19) Q=ZW q (20) V=ZW v (twenty one)

[0067] in, and With the help of K and Q, we can get the attention matrix It can be expressed as follows:

[0068] in, is a scaling factor. Therefore, the output of the rth component of the attention mechanism can be expressed as the weighted sum of all inputs:

[0069] in, represents the r-th output, which is calculated by adaptively focusing on the input according to the attention score AM[r,j]. Finally, the overall attention map can be expressed as:

[0070] in,

[0071] Based on the above, the present invention proposes an MHSA module for CGM construction. Specifically, given the output of the LRDC module at the (L-1)th layer First input to the normalization layer, and then through linear transformation to obtain the query vector Key Vector Sum vector Then, the single-head self-attention map is obtained by the following formula:

[0072] According to formula (25), it can be extended to a multi-head self-attention learning mechanism. First, the self-attention subgraph learned in the i-th subspace can be calculated as follows:

[0073] in, yes Mapping in the i-th learning subspace. W i is a trainable transformation matrix expressed as:

[0074] (28) Mapped to different self-attention learning subspaces, forming a self-attention map of I heads. Then, the self-attention submaps corresponding to the I transformed subspaces are spliced ​​using the following formula:

[0075] Among them, W t is a trainable matrix, I is the number of parallel self-attention learning subspaces. The obtained self-attention map For high frequency subgraphs rL-1 Refinement:

[0076] in, Represents pixel-level multiplication. Through the above steps, the high-frequency components in the (L-1) layer of the Laplacian pyramid are completed. Pixel-level refinement.

[0077] For the remaining frequency components R=[r0,r1,r2,...,r L-2 ]'s refinement will Extended to the corresponding self-attention map The refined results of the residual frequency components are obtained through upsampling operation, LRDC and MHCCA modules in Then, using the reversibility of the Laplace pyramid, we use the reconstructed and refined Get your target CGM.

[0078] 3) MHCCA module: The MHCCA module uses the cross-covariance attention mechanism to calculate the attention score along the feature dimension rather than the spatial (or token) dimension. It is a transposed version of self-attention (22) and can be expressed as: ψ cc =V·AM cc (Q,K) (33)

[0079] Among them, AM cc (Q,K),ψ cc are the cross-covariance attention matrix and graph, respectively. ε is a learnable scaling parameter.

[0080] IV. Implementation Effect

[0081] In order to enable those skilled in the art to better understand the solution of the present invention, the following provides a comparison of the estimated performance results of the channel knowledge map reconstruction method in this embodiment and the existing method under several specific system configurations.

[0082] First, the impact of the number of attention heads on the performance of this example implementation is presented when the number of Laplacian pyramid layers is 3. When the number of attention heads increases from 1 to 5, the proposed LPCGMN shows good robustness in reducing prediction error with increasing the number of attention heads I. Specifically, as shown in Figure 5, the reconstruction gain of the LPCGMN with I = 5 is improved by 1.72dB compared to the LPCGMN with I = 1. However, when I is set too large, such as greater than 5, the prediction error of the proposed LPCGMN does not decrease, but instead increases slightly. Compared with the LPCGMN with I = 5, the performance gain of the LPCGMN with I = 6 and I = 7 is reduced by 0.66dB and 0.22dB, respectively. Analysis suggests that this is due to the insufficient size of the tested image. However, when the number of learned subspaces is too large, more subspaces lose their independence from each other, resulting in overlap between them, which affects the effective learning of the LPCGMN and leads to stagnation.

[0083] Next, a schematic diagram comparing the complexity performance of the LPCGMN in this embodiment with existing methods is given. The comparison methods are the Unet method proposed in the document “U-Net: Convolutional networks for biomedical image segmentation,” in Proc. Int. Conf. Med. Image Comput. Comput.-Assist. Interv. (MICCAI), Munich, Germany, Oct. 2015, pp. 234–241.” and the Wnet method in the document “RadioUnet: Fast radio map estimation with convolutional neural networks,” IEEE Trans. Wirel. Commun., vol. 20, no. 6, pp. 4001–4015, Jun. 2021.” Figure 6 shows the performance comparison of the LPCGMN in this embodiment with the Unet and Wnet methods on the same dataset under the considered wireless communication system. The designed performance indicators include the time spent in each training cycle, time complexity (FLOPs), storage complexity (Parameters), and prediction error (NMSE, Normalized Mean Square Error). The training epochs for LPCGMN, Unet, and Wnet were all set to 50. As shown in Figure 6, Wnet required a longer training time per epoch, approximately 16.68 minutes. The proposed LPCGMN (L=5, I=1) had the shortest training time, approximately 1.13 minutes. In terms of FLOPs, Wnet still achieved the highest performance, approximately 94.16 Giga FLOPs. LPCGMN (L=3, I=1) and LPCGMN (L=5, I=1) achieved performances of 6.16 and 3.86 Giga FLOPs, respectively. Furthermore, when increasing the number of LP layers to reduce FLOPs, the proposed LPCGMN maintained stable performance. In terms of model parameters, the lightest model was LPCGMN (L=3, I=1), with approximately 10.12 million parameters, while Wnet had a larger number of parameters, approximately 26.54 million. In terms of NMSE, the proposed LPCGMN (L=3, I=5) has a lower reconstruction error of approximately 0.0064, which indicates that compared with the Unet and Wnet methods, the LPCGMN in this embodiment can accurately construct the CGM at a faster speed.

[0084] Based on the same inventive concept, an embodiment of the present invention discloses a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded into the processor, the method for rapidly constructing a channel knowledge map applicable to a base station or a user terminal is implemented.

[0085] In a specific implementation, the device includes a processor, a communication bus, a memory, and a communication interface. The processor can be a general-purpose central processing unit (CPU), a microprocessor, an application-specific integrated circuit (ASIC), or one or more integrated circuits used to control the execution of the program of the present invention. The communication bus can include a path for transmitting information between the above components. The communication interface uses any device such as a transceiver for communicating with other devices or communication networks. The memory can be a read-only memory (ROM) or other type of static storage device that can store static information and instructions, a random access memory (RAM) or other type of dynamic storage device that can store information and instructions, an electrically erasable programmable read-only memory (EEPROM), a compact disc read-only memory (CD-ROM) or other optical storage, a disk storage medium or other magnetic storage device, or any other medium that can be used to carry or store desired program code in the form of instructions or data structures and can be accessed by a computer, but is not limited to these. The memory can be independent and connected to the processor via a bus. The memory can also be integrated with the processor.

[0086] The memory is used to store application code for executing the solution of the present invention, and the execution is controlled by the processor. The processor is used to execute the application code stored in the memory, thereby implementing the channel estimation method provided by the above embodiment. The processor may include one or more CPUs, or multiple processors, each of which may be a single-core processor or a multi-core processor. The processor here may refer to one or more devices, circuits, and / or processing cores for processing data (e.g., computer program instructions).

[0087] Based on the same inventive concept, an embodiment of the present invention discloses a wireless communication system comprising a base station and multiple user terminals. The base station or user terminal is provided with: an information processing module for acquiring environmental information of a target area and converting the environmental information into a 2D image format through a spatial discretization method and pixel-level grayscale conversion, thereby converting the channel knowledge map construction problem into an image restoration task; a frequency band decomposition module for decomposing the environmental map into different frequency bands using a Laplacian pyramid to obtain different frequency components; a network model construction module for constructing a channel knowledge map reconstruction network model; the network model designs different subnetworks for feature extraction for different frequency components of the environmental map. After the different frequency components are output by their respective subnetworks, the channel knowledge map is reconstructed using the inverse operation of the Laplacian pyramid; a model training module for training the constructed network model using an end-to-end supervised training method based on a training set containing the environmental map and its corresponding channel knowledge map to obtain the optimal model parameters; and a prediction module for using the trained model to predict the channel knowledge information of user equipment at a specific location in the target area. The specific implementation of each module is described in the aforementioned method embodiment and will not be repeated here.

[0088] In the embodiments provided herein, it should be understood that the disclosed methods may be implemented in other ways without departing from the spirit and scope of the present application. The present embodiments are merely illustrative examples and should not be construed as limiting. The specific details provided herein should not limit the purpose of the present application. For example, some features may be omitted or not implemented.

[0089] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A method for rapidly constructing a wireless channel knowledge map, characterized in that: The following steps are involved: Obtain the environmental information of the target area and convert it into a 2D image format through spatial discretization method and pixel-level grayscale conversion, thereby converting the channel knowledge map construction problem into an image restoration task; Use the Laplacian pyramid to decompose the environment map into different frequency bands to obtain different frequency components; Constructing a channel knowledge map reconstruction network model; the network model designs different sub-networks for different frequency components of the environment map to extract features, and after the different frequency components are output by their respective sub-networks, the channel knowledge map is reconstructed by using the inverse operation of the Laplace pyramid; Given a training set containing an environment map and its corresponding channel knowledge map, the constructed network model is trained using an end-to-end supervised training method to obtain the optimal parameters of the model; The trained model is used to predict the channel knowledge information of user devices at specific locations in the target area.

2. A method for rapidly constructing a wireless channel knowledge map according to claim 1, characterized in that: The transformation of the channel knowledge map construction problem to the image restoration problem is to treat each specific position as a pixel point of the image matrix, and the channel knowledge corresponding to the specific position as the pixel value of the corresponding pixel point, and transform the channel knowledge estimation problem of each specific position into a pixel restoration problem between images.

3. A method for rapidly constructing a wireless channel knowledge map according to claim 1, characterized in that: The spatial discretization is to grid the environmental information in the spatial dimension, and divide the user equipment position into the nearest grid according to the minimum distance criterion.

4. The method for rapidly constructing a wireless channel knowledge map according to claim 1, characterized in that: The grayscale conversion is implemented at the pixel level, and the channel knowledge value is converted into an interval range from 0 to 1 using minimum-maximum normalization.

5. The method for rapidly constructing a wireless channel knowledge map according to claim 1, characterized in that: For a pair of environment maps and channel knowledge maps, the difference in low-frequency components is more obvious than that in high-frequency components; the sub-network corresponding to the low-frequency components has more feature extraction levels than the sub-network corresponding to the high-frequency components.

6. A method for rapidly constructing a wireless channel knowledge map according to claim 1, characterized in that: The network model mainly includes: Reconstruction of low-frequency components of layer L: low-frequency environment sub-image I L First, it undergoes a point convolution to complete the dimensional expansion in the depth direction; then, it is input into multiple stacked lightweight residual dilated convolution LRDC modules to deepen the feature extraction network and obtain the fusion of multi-scale features; then, it is input into the multi-head self-attention MHSA module or the multi-head cross-covariance self-attention MHCCA module to encode the global random structure information into the local features of the spatial dimension and learn a richer hierarchical feature representation; finally, the number of color channels in the feature map is reduced to the original size to obtain the output result Refined processing of high-frequency components at layer L-1: The low-frequency components and their reconstruction results are fused to gradually guide the refined processing of high-frequency components at different levels to obtain components of different frequencies for reconstructing the channel knowledge map; first, the low-frequency environment sub-map I L and the predicted channel knowledge subgraph Perform bilinear upsampling respectively; then, I L , and the high-frequency component r of the L-1 layer L-1 The L-1th layer high-frequency component r is obtained by splicing and inputting into multiple superimposed LRDC modules; then, the extracted multi-scale feature map is input into the MHSA module or the MHCCA module; finally, the L-1th layer high-frequency component r is obtained. L-1 The refined result and the attention map used to refine the high-frequency components of the L-2 layer; For the remaining frequency components R = [r0, ..., r L-2 ], through iterative upsampling, concatenation operation, LRDC module, and MHSA module or MHCCA module, the high-frequency components of all levels of the Laplacian pyramid are gradually refined, and the prediction results of different components are obtained.

7. A method for rapidly constructing a wireless channel knowledge map according to claim 6, characterized in that: The LRDC module consists of a depth-separable atrous convolution layer, an instance normalization layer, a point convolution layer and a GELU activation function layer, which is used for feature extraction and representation; the MHSA module consists of a normalization layer, a multi-head self-attention layer, a point convolution layer and a GELU activation function layer; the MHCCA module consists of a normalization layer, a multi-head cross-covariance attention layer, a point convolution layer and a GELU activation function layer, which is used to encode global environmental information.

8. A method for rapidly constructing a wireless channel knowledge map according to claim 1, characterized in that: The cost function of the network model training is expressed as: Among them, K is the total number of samples in the training set, is the reconstructed channel knowledge map, It is a real channel knowledge map. represents the constructed network model, Θ is The set of parameters to be learned in represents the environment map, e represents the propagation environment, x n represents the user device location, and ||·||2 represents the Euclidean norm.

9. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the computer program is loaded into a processor, a method for quickly constructing a wireless channel knowledge map according to any one of claims 1 to 8 is implemented.

10. A wireless communication system, comprising a base station and a plurality of user terminals, characterized in that: The base station or user terminal is provided with: The information processing module is used to obtain the environmental information of the target area and convert the environmental information into a 2D image format through spatial discretization method and pixel-level grayscale conversion, thereby converting the channel knowledge map construction problem into an image restoration task; The frequency band decomposition module is used to decompose the environment map into different frequency bands using the Laplacian pyramid to obtain different frequency components; A network model construction module is used to construct a channel knowledge map reconstruction network model; the network model designs different sub-networks for different frequency components of the environment map to extract features, and after the different frequency components are output by their respective sub-networks, the channel knowledge map is reconstructed by means of an inverse operation of the Laplace pyramid; The model training module is used to train the constructed network model using an end-to-end supervised training method given a training set including an environment map and its corresponding channel knowledge map to obtain the optimal parameters of the model; And, a prediction module is used to use the trained model to predict the channel knowledge information of the user equipment at a specific location in the target area.

Images