Channel power spectrum estimation method and device, electronic equipment and storage medium

Abstract

The embodiment of the invention provides a channel power spectrum estimation method and device, electronic equipment and a storage medium, and the method comprises the steps: obtaining a grid position of a to-be-estimated grid in a target region, and obtaining a cell position and a cell antenna parameter of at least one communication cell corresponding to the grid position, the live-action statistical channel model is obtained by training reference signal receiving power between a plurality of grids in the target area and a communication cell in advance; inputting the cell position and the cell antenna parameter into a cell feature encoder for feature encoding processing to obtain a cell feature code; inputting the grid position and the cell position into a grid feature encoder for feature encoding processing to obtain a grid feature code; and inputting the cell feature code and the grid feature code into an angle power spectrum estimation network for power spectrum estimation to obtain an estimated channel power spectrum between the grid to be estimated and the communication cell, thereby ensuring the accuracy of channel estimation and improving the channel estimation efficiency in a wireless communication system.

Description

Technical Field

[0001] This application relates to the technical field of communication data processing, and in particular, to a channel power spectrum estimation method, apparatus, electronic device, and storage medium. Background Art

[0002] The performance of a wireless communication system is mainly affected by the wireless channel. Different from the usually static and predictable wired channel, the wireless channel has dynamics and unpredictability, which often makes the accurate analysis of the wireless communication system difficult. Taking network optimization as an example, it can significantly improve the quality of service of users in the network and increase the throughput, but there is still a large gap between the results of network optimization and the actual application scenarios. One important reason for this is that the complex wireless channels in the existing network cannot be accurately characterized and estimated, resulting in the inability to effectively improve the network performance.

[0003] In the related art, for channel estimation of a complex wireless communication system, in order to improve the accuracy of channel estimation, usually according to the complex high-precision environmental information in the wireless communication system, and then perform deterministic modeling to use these high-precision environmental information for accurate channel estimation. However, since the environmental information in the actual wireless communication system is usually time-varying, if it is necessary to obtain complex high-precision environmental information in real time and rely on these real-time high-precision environmental information for real-time channel estimation, the computational complexity is very high, resulting in a low channel estimation efficiency in the existing channel estimation methods. Summary of the Invention

[0004] Embodiments of this application provide a channel power spectrum estimation method, apparatus, electronic device, and storage medium, which can improve the channel estimation efficiency in a wireless communication system.

[0005] To achieve the above object, a first aspect of the embodiments of this application proposes a channel power spectrum estimation method, which is executed by a virtualized statistical channel model. The virtualized statistical channel model at least includes: a cell feature encoder, a grid feature encoder, and an angular power spectrum estimation network. The method includes: Obtaining the grid position of the grid to be estimated in the target area, and obtaining the cell position and cell antenna parameters of at least one communication cell corresponding to the grid position. The virtualized statistical channel model is pre-trained by the reference signal received power between multiple grids in the target area and the communication cell; Inputting the cell position and the cell antenna parameters into the cell feature encoder for feature encoding processing to obtain a cell feature encoding; Inputting the grid position and the cell position into the grid feature encoder for feature encoding processing to obtain a grid feature encoding; Input the cell feature encoding and the grid feature encoding into the angular power spectrum estimation network for power spectrum estimation to obtain the estimated channel power spectrum between the grid to be estimated and the communication cell.

[0006] In some embodiments, the virtualized statistical channel model further includes a transformer module. The step of inputting the cell feature encoding and the grid feature encoding into the angular power spectrum estimation network for power spectrum estimation to obtain the estimated channel power spectrum between the grid to be estimated and the communication cell includes: Perform exclusive OR processing on the cell feature encoding and the grid feature encoding, and then superimpose the position embedding of the grid position to obtain a feature processing unit; Input the feature processing unit into the transformer module for attention processing to obtain estimated processing data; Input the estimated processing data into the angular power spectrum estimation network for power spectrum estimation to obtain the estimated channel power spectrum.

[0007] In some embodiments, the transformer module includes a multi-head attention module. The step of inputting the feature processing unit into the transformer module for attention processing to obtain estimated processing data includes: Obtain other feature processing units between other grids in the target area and the communication cell, and combine the other feature processing units and the feature processing unit to obtain a feature unit matrix; Input the feature unit matrix into the multi-head attention module for multi-head attention processing to obtain the estimated processing data.

[0008] In some embodiments, the multi-head attention module includes multiple self-attention processing modules. The step of inputting the feature unit matrix into the multi-head attention module for multi-head attention processing to obtain the estimated processing data includes: In the multi-head attention module, based on the ratio of the data dimension of the feature unit matrix to the number of attention heads, obtain the division dimension value of each attention head; Generate a projection parameter matrix for each attention head based on the division dimension value; In each self-attention processing module, generate a matrix query vector, a matrix key vector, and a matrix value vector based on the feature unit matrix, and perform attention processing on the matrix query vector, the matrix key vector, and the matrix value vector based on the corresponding projection parameter matrix and linear transformation matrix to obtain sub-estimated processing data; Perform multi-head attention processing on multiple sub-estimated processing data to obtain the estimated processing data.

[0009] In some embodiments, inputting the estimated processing data into the angular power spectrum estimation network for power spectrum estimation to obtain the estimated channel power spectrum includes: In the angular power spectrum estimation network, linearly process, perform activation function processing, layer normalization processing, and regularization processing on the estimated processing data in sequence to obtain preliminarily processed data; Perform residual connection enhancement processing and feature transfer processing on the preliminary processing result in sequence to obtain enhanced processing data; Perform non - negative activation processing on the enhanced processing data to obtain activated processing data; Perform estimation processing based on the activated processing data to obtain the estimated channel power spectrum.

[0010] In some embodiments, the training steps of the realistic statistical channel model include: Obtain the channel impulse response function and the precoding matrix between each antenna and each grid in the communication cell; Accumulate the product of all the channel impulse response functions and the precoding matrix, perform a square process, and then multiply by the transmit power of the communication cell to obtain the reference signal received power function for each grid; Perform time expectation processing on the reference signal received power function to obtain the expected reference signal received power function, and the expected reference signal received power function includes the coefficient matrix between the grid and the communication cell; Generate a channel power spectrum estimation model based on the coefficient matrix, the reference signal received power function, and the channel power spectrum estimation function between each grid and the communication cell; Perform multiple iterative updates on the realistic statistical channel model based on the channel power spectrum estimation model and multiple reference signal received powers.

[0011] In some embodiments, performing multiple iterative updates on the realistic statistical channel model based on the channel power spectrum estimation model includes: Obtain a sparse penalty weight, and obtain a penalty channel power spectrum function based on the product of the sparse penalty weight and the channel power spectrum estimation function; Generate a loss function based on the accumulated value of the channel power spectrum estimation model and the penalty channel power spectrum function; Perform multiple iterative updates on the realistic statistical channel model based on the loss function and multiple reference signal received powers.

[0012] To achieve the above object, a second aspect of the embodiments of the present application proposes a channel power spectrum estimation device, which is executed by a virtualized statistical channel model. The virtualized statistical channel model at least includes: a cell feature encoder, a grid feature encoder, and an angular power spectrum estimation network. The device includes: A data acquisition module, configured to acquire the grid position of the grid to be estimated in the target area, and acquire the cell position and cell antenna parameters of at least one communication cell corresponding to the grid position. The virtualized statistical channel model is pre-trained by the reference signal received power between multiple grids in the target area and the communication cell; A cell feature processing module, configured to input the cell position and the cell antenna parameters into the cell feature encoder for feature encoding processing to obtain a cell feature code; A grid feature processing module, configured to input the grid position and the cell position into the grid feature encoder for feature encoding processing to obtain a grid feature code; A channel power spectrum estimation module, configured to input the cell feature code and the grid feature code into the angular power spectrum estimation network for power spectrum estimation to obtain the estimated channel power spectrum between the grid to be estimated and the communication cell.

[0013] To achieve the above object, a third aspect of the embodiments of the present application proposes an electronic device. The electronic device includes a memory and a processor. The memory stores a computer program. When the processor executes the computer program, the channel power spectrum estimation method described in the first aspect is implemented.

[0014] To achieve the above object, a fourth aspect of the embodiments of the present application proposes a storage medium. The storage medium is a computer-readable storage medium. The storage medium stores a computer program. When the computer program is executed by a processor, the channel power spectrum estimation method described in the first aspect is implemented.

[0015] The channel power spectrum estimation method, device, electronic device, and storage medium proposed in the embodiments of this application are executed by a virtualized statistical channel model. The virtualized statistical channel model includes at least: a cell feature encoder, a grid feature encoder, and an angular power spectrum estimation network. The method includes: First, obtain the grid position of the grid to be estimated in the target area, and obtain the cell position and cell antenna parameters of at least one communication cell corresponding to the grid position. The virtualized statistical channel model is pre-trained by the reference signal received power between multiple grids and communication cells in the target area; Then, input the cell position and cell antenna parameters into the cell feature encoder for feature encoding processing to obtain cell feature encoding; Next, input the grid position and cell position into the grid feature encoder for feature encoding processing to obtain grid feature encoding; Finally, input the cell feature encoding and grid feature encoding into the angular power spectrum estimation network for power spectrum estimation to obtain the estimated channel power spectrum between the grid to be estimated and the communication cell. In the embodiments of this application, a virtualized statistical channel model is trained by using the measured data of the reference signal received power with low acquisition cost between the target area and the communication cell in advance, and adopting a method of fusing physical model guidance and deep learning, avoiding the real-time, complex, high-dimensional environmental information acquisition and high-computation deterministic modeling required by traditional high-precision methods, so as to reduce the training complexity of the estimation model for channel power spectrum estimation and improve the training efficiency. And because the model directly learns from the measured data of the reference signal received power that reflects the influence of the real environment, it is possible to generate a channel power spectrum estimation result that is closer to the actual situation and has more "virtualized" characteristics; In actual channel power spectrum estimation, use the grid position of the grid to be estimated that can be easily obtained in real time, the cell position and cell antenna parameters in the communication cell, and the pre-obtained virtualized statistical channel model for efficient feature extraction, thus avoiding the need for real-time acquisition of complex and time-varying high-precision environmental information, significantly reducing the computational complexity, and efficiently obtaining the real-time estimated channel power spectrum in the wireless communication system. Furthermore, while ensuring the accuracy of channel estimation, the channel estimation efficiency in the wireless communication system is effectively improved.

[0016] Other features and advantages of this application will be described in the following specification, and part of them will become obvious from the specification, or be understood by implementing this application. The purpose and other advantages of this application can be realized and obtained through the structures specifically pointed out in the specification, claims, and drawings. Brief Description of the Drawings

[0017] Figure 1 It is a flowchart of a channel power spectrum estimation method provided by an embodiment of this application.

[0018] Figure 2 It is a training flowchart of the virtualized statistical channel model provided by another embodiment of this application.

[0019] Figure 3 It is a schematic structural diagram of a scene-based statistical channel model provided by another embodiment of the present application.

[0020] Figure 4 It is Figure 2 The flowchart of step 205 in

[0021] Figure 5 It is a schematic structural diagram of a cell feature encoder and a grid feature encoder provided by another embodiment of the present application.

[0022] Figure 6 It is Figure 1 The flowchart of step 104 in

[0023] Figure 7 It is Figure 6 The flowchart of step 602 in

[0024] Figure 8 It is a schematic structural diagram of a multi-head attention module provided by another embodiment of the present application.

[0025] Figure 9 It is Figure 7 The flowchart of step 702 in

[0026] Figure 10 It is a schematic structural diagram of a self-attention processing module provided by another embodiment of the present application.

[0027] Figure 11 It is Figure 6 The flowchart of step 603 in

[0028] Figure 12 It is a schematic structural diagram of an angular power spectrum estimation network provided by an embodiment of the present application.

[0029] Figure 13 It is a schematic conceptual diagram of a channel power spectrum estimation provided by an embodiment of the present application.

[0030] Figure 14 It is a schematic simulation diagram of the loss decline curve of a scene-based statistical channel model provided by an embodiment of the present application.

[0031] Figure 15 It is a schematic performance simulation diagram of a channel power spectrum estimation method provided by an embodiment of the present application.

[0032] Figure 16 It is a schematic structural diagram of a channel power spectrum estimation device provided by an embodiment of the present application.

[0033] Figure 17It is a schematic diagram of the hardware structure of an electronic device provided by an embodiment of the present application. Detailed implementation manners

[0034] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, but not to limit the present application.

[0035] It should be noted that although the functional modules are divided in the device schematic diagram and the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in a different order from the module division in the device or the flowchart.

[0036] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the technical field to which this application belongs. The terms used herein are only for the purpose of describing the embodiments of this application and are not intended to limit this application.

[0037] The performance of a wireless communication system is mainly affected by the wireless channel. Different from the usually static and predictable wired channel, the wireless channel has dynamics and unpredictability, which makes the accurate analysis of the wireless communication system often difficult. Taking network optimization as an example, it can significantly improve the quality of service of users in the network and increase the throughput, but there is still a large gap between the result of network optimization and the actual application scenario. An important reason for this is that the complex wireless channels in the existing network cannot be accurately characterized and estimated, resulting in the inability to effectively improve the network performance.

[0038] In the related art, for channel estimation of a complex wireless communication system, in order to improve the accuracy of channel estimation, usually according to the complex high-precision environmental information in the wireless communication system, and then deterministic modeling is performed to use these high-precision environmental information for accurate channel estimation. However, since the environmental information in the actual wireless communication system is usually time-varying, if it is necessary to obtain complex high-precision environmental information in real time and rely on these real-time high-precision environmental information for real-time channel estimation, the computational complexity is very high, resulting in a relatively low channel estimation efficiency in the existing channel estimation methods.

[0039] To improve the channel estimation efficiency in a wireless communication system, embodiments of the present application obtain a measured reference signal received power data with low acquisition cost between a target area and a communication cell in advance, and train a real-scene statistical channel model by combining physical model guidance and deep learning, avoiding the real-time, complex, high-dimensional environmental information acquisition and high computational deterministic modeling required by traditional high-precision methods, so as to reduce the training complexity of the estimation model for channel power spectrum estimation and improve the training efficiency. Since the model directly learns from the measured reference signal received power data reflecting the influence of the real environment, it is possible to generate a channel power spectrum estimation result that is closer to the actual situation and has more "real-scene" characteristics. In actual channel power spectrum estimation, efficient feature extraction is performed using the grid position of the grid to be estimated that can be easily obtained in real time, the cell position and cell antenna parameters in the communication cell, and the previously obtained real-scene statistical channel model, thereby avoiding the need for real-time acquisition of complex and time-varying high-precision environmental information, significantly reducing the computational complexity, and efficiently obtaining the real-time estimated channel power spectrum in the wireless communication system. Furthermore, while ensuring the accuracy of channel estimation, the channel estimation efficiency in the wireless communication system is effectively improved.

[0040] The following describes the channel power spectrum estimation method, apparatus, electronic device, and storage medium provided by embodiments of the present application. The channel power spectrum estimation method provided in embodiments of the present application can be applied to any server or computing processor with computing resources, etc.

[0041] The following will specifically describe the channel power spectrum estimation method in embodiments of the present application. Refer to Figure 1 , which is an optional flowchart of the channel power spectrum estimation method provided by embodiments of the present application. Figure 1 The method in Figure 1 may include but is not limited to steps 101 to 104. At the same time, it can be understood that the order of steps 101 to 104 in this embodiment is not specifically limited, and the step order can be adjusted according to actual needs, or some steps can be reduced or added.

[0042] Step 101: Obtain the grid position of the grid to be estimated in the target area, and obtain the cell position and cell antenna parameters of at least one communication cell corresponding to the grid position.

[0043] The following details step 101.

[0044] In some embodiments, for a wireless communication system including a target area and at least one communication cell, the target area includes multiple grids, and the channel from each communication cell to each grid can be characterized by large-scale fading (path loss, shadow fading) and small-scale fading. Among them, each communication cell (such as the The antenna array of a communication cell includes antennas.

[0045] The target area can be a hot spot area or a specific coverage area of a city. Then, within the target area, one or more grids are selected as the grids to be estimated (i.e., any one grid or multiple grids among the multiple grids of the target area, such as the th grid), and the grid position can be represented by longitude and latitude coordinates, for example, (39.9042°N, 116.4074°E).

[0046] Next, in the wireless communication system, determine the communication cell associated with the grid to be estimated, or the communication cell that the user needs to determine. This can be obtained through querying the geographical location database or obtaining network planning information. For each associated communication cell, its cell location (which can also be represented by longitude and latitude coordinates) and cell antenna parameters need to be obtained. The cell antenna parameters include parameters such as antenna gain, antenna angle, and frequency point. These parameters are usually stored in the network management device of the wireless communication system and can be obtained through interface calls or data exports. These data can be easily obtained without the need for complex acquisition devices or complex data processing means.

[0047] After obtaining the grid position of the grid to be estimated, the cell location of the communication cell, and the cell antenna parameters, use the real-scene statistical channel model pre-trained by the reference signal received power between multiple grids and communication cells in the target area to perform channel power spectrum estimation, so as to quickly obtain the estimated channel power spectrum between the grid to be estimated and the communication cell, thereby effectively improving the estimation efficiency of channel estimation.

[0048] First, the following describes how to train the real-scene statistical channel model.

[0049] Referring to Figure 2 , the training steps of the real-scene statistical channel model include the following steps 201 to 205.

[0050] Step 201: Obtain the channel impulse response function and precoding matrix between each antenna in the communication cell and each grid.

[0051] Step 202: Accumulate the product of all channel impulse response functions and precoding matrices, perform a square process, and then multiply by the transmission power of the communication cell to obtain the reference signal received power function for each grid.

[0052] Step 203: Perform time expectation processing on the reference signal received power function to obtain the expected reference signal received power function.

[0053] Step 204: Generate a channel power spectrum estimation model based on the coefficient matrix, the reference signal received power function, and the channel power spectrum estimation function between each grid and the communication cell.

[0054] Step 205: Perform multiple iterative updates on the virtualized statistical channel model based on the channel power spectrum estimation model and multiple reference signal received powers.

[0055] The following provides a detailed description of Steps 201 to 205.

[0056] In some embodiments, for the above-mentioned wireless communication system, the channel impulse response function between each antenna (with its antenna position being ) in the th communication cell and the th grid in the target area at

[0057] (1) where and respectively correspond to the total number of tangent points in the angular domain of the vertical plane and the horizontal plane, is the antenna gain of a certain antenna (with its antenna position being ) in the th communication cell, is the elevation angle of incidence of the th communication cell, is the azimuth angle of incidence of the th communication cell, and are the spacings between adjacent antennas in the corresponding directions, is the angular phase error that follows a uniform distribution in the interval , is the antenna - to - antenna phase error that follows a Gaussian distribution with a mean of 0 and a variance of , is the wavelength of the electromagnetic wave, is the channel loss between a certain antenna (with its antenna position being ) in the th communication cell and the th grid in the target area at moment, and this channel loss follows a log - normal distribution as shown in the following formula (2).

[0058] (2) where and are respectively the channel loss between a certain antenna (with its antenna position being ) and the mean and standard deviation of the log-normal distribution of the path loss between the th grid in the target area.

[0059] In large-scale antennas, since the dimension of the channel impulse response is large, in order to reduce complexity, the reference signal received power rsrp of multiple beams is considered in this embodiment. Assume that the th beam in the antenna array of the th communication cell corresponds to a precoding matrix as shown in the following formula (3).

[0060] (3) where is the phase of the th beam in the th communication cell. Then, accumulate the product of all channel impulse response functions (1) and precoding matrices (3), perform a square operation, and then multiply by the transmit power of the communication cell to obtain the reference signal received power function for each grid. For example, the reference signal received power function between the th beam in the th communication cell and the th grid is as shown in the following formula (4).

[0061] (4) Based on formula (1), further expand the reference signal received power function as shown in the following formula (5).

[0062] (5) Next, perform a time expectation process on the reference signal received power function (5) using the time parameter t to obtain the expected reference signal received power function as shown in the following formula (6).

[0063] (6) where the expected reference signal received power function includes the coefficient matrix between the grid and the communication cell, as shown in the following formula (7).

[0064] (7) (8) So far, a statistical relationship between the reference signal received power and the angular power spectrum has been established in this embodiment. For a single-cell single-grid, this relationship can be as shown in the following formula (9).

[0065] , (9) Among them, the coefficient matrix has a main lobe and side lobes. The paths between the antenna array and the grid can only exist at the angles corresponding to the main lobe and side lobes. The angle power spectrum to be estimated is a sparse vector, and the non-zero elements represent the power of the multipath channel and the corresponding angles.

[0066] The goal of the realistic statistical channel modeling is to estimate through collecting the measured data of the received power of the reference signal and according to its statistical relationship with the channel angle power spectrum . characterizes the statistical distribution of the channel power in the angle domain, reflects the multipath topology of the realistic environment, and can be used for the performance prediction after the network parameters are optimized.

[0067] Due to the large number of communication cells and the complex coupling relationship in the optimization of the actual network corresponding to the wireless communication system, it is necessary to consider the channel modeling in a large range. Specifically, the goal in the embodiment of this application is the large-scale realistic statistical channel modeling for multiple communication cells and multiple grids. Therefore, the channel angle power spectrum estimation is expressed as a function parameterized by , then represents the channel angle power spectrum of the th communication cell at the th grid; then based on the coefficient matrix , the reference signal received power function (5), and the channel power spectrum estimation function between each grid and the communication cell, a channel power spectrum estimation model is generated as shown in the following formula (10).

[0068] (10) Among them, is the total number of grid cells in the target area, P is the maximum number of paths of the multipath channel, and the channel power spectrum is non-negative.

[0069] For the sparse recovery problem of the channel power spectrum estimation model (10), the traditional optimization algorithms use greedy or iterative methods to solve it, and cannot effectively utilize the environmental information. In this embodiment, relying on the powerful learning ability of the deep neural network, a realistic statistical channel model based on the Transformer is designed to learn from the data and establish an implicit mapping from the environmental information to the complex non-linear channel angle power spectrum.

[0070] Referring to Figure 3 , it is a schematic structural diagram of a realistic statistical channel model provided by the embodiment of this application. As Figure 3As shown in [FIGURE], the virtualized statistical channel model consists of a cell feature encoder, a grid feature encoder, a transformer, and an angular power spectrum estimation network. The cell feature encoder receives the cell location and cell antenna parameters as inputs, and the grid feature encoder receives the grid location and cell location as inputs, respectively, for feature encoding. The encoded features are added to the geospatial location embedding and then input into the transformer module for attention processing. The output of the transformer module is finally input into the angular power spectrum estimation network, which contains structures such as layer normalization, feed-forward network, multi-head self-attention mechanism, and residual connection inside, and finally outputs the estimated channel power spectrum. Among them, "+" represents feature addition, and "C" represents feature concatenation.

[0071] After constructing the virtualized statistical channel model, the virtualized statistical channel model will be iteratively updated multiple times based on the channel power spectrum estimation model (10) and the reference signal received power between multiple grids and communication cells obtained from multiple actual measurements in the target area, as described below.

[0072] Refer to Figure 4 , based on the channel power spectrum estimation model and multiple reference signal received powers, the virtualized statistical channel model is iteratively updated multiple times, including the following steps 401 to step 403.

[0073] Step 401: Obtain the sparse penalty weight, and based on the product of the sparse penalty weight and the channel power spectrum estimation function, obtain the penalty channel power spectrum function.

[0074] Step 402: Generate a loss function based on the cumulative value of the channel power spectrum estimation model and the penalty channel power spectrum function.

[0075] Step 403: Iteratively update the virtualized statistical channel model multiple times based on the loss function and multiple reference signal received powers.

[0076] The following provides a detailed description of steps 401 to 403.

[0077] In some embodiments, since the propagation environments of different communication cells are different, when constructing the multi-grid sequence, each grid is paired with the same communication cell. To effectively train the entire virtualized statistical channel model, this embodiment designs a model training framework based on self-supervised learning.

[0078] Based on the pre-set sparse penalty weight and the product of the norm of the channel power spectrum estimation function, obtain the penalty channel power spectrum function , and then generate a loss function based on the cumulative value of the objective function in the channel power spectrum estimation model (10) and the penalty channel power spectrum function as shown in the following formula (11).

[0079] (11) The loss function includes a reconstruction loss and a sparsity penalty, and the reconstruction objective is the maximum value of the multi-beam reference signal reception power.

[0080] After that, based on this loss function and the reference signal reception powers between multiple grids in the target area obtained through multiple measurements and the communication cell, the virtualized statistical channel model is iteratively updated multiple times. In the virtualized statistical channel model shown as Figure 3 below, the module parameters in each iterative update include the relevant parameters in the cell feature encoder, grid feature encoder, transformer, and angular power spectrum estimation network.

[0081] Once the training is completed, the model can estimate the channel angular power spectrum of the grids of a specified cell according to the given input, and then predict the maximum value of the reference signal reception power after the network parameters in the wireless communication system are adjusted in combination with formula (6), so as to accurately perform channel estimation on this wireless communication system.

[0082] Through the above steps 201 to 205, and steps 401 to 403, starting from basic physical quantities such as channel impulse response, precoding, and transmit power, a clear mathematical relationship between the expected reference signal reception power and the underlying channel statistical characteristics (finally reflected as the angular power spectrum and the coefficient matrix) is derived. This provides a solid physical basis and interpretable basis for subsequent model training, avoids the drawback of the pure black-box model lacking physical meaning, constructs a learning objective using this physical relationship, and iteratively trains the deep learning model with the actually measured and easily obtained reference signal reception power data, enabling the model to learn the complex propagation characteristics in a specific "virtual" environment, overcoming the problems of poor generalization ability of traditional statistical models or large computational complexity of deterministic models that rely on high-precision environmental information, realizing the "virtualization" and high efficiency of the model, and by introducing a sparsity penalty term in the loss function, using the prior knowledge that the angular power spectrum of the wireless channel usually has sparsity, guiding the model to learn a more physically realistic and concise and effective power spectrum representation, which not only helps to improve the generalization ability and estimation accuracy of the model, but also can improve the robustness of the model to a certain extent, so as to obtain an efficient, accurate and practical channel power spectrum estimation model that combines physical principles, measured data and prior knowledge.

[0083] Since in the embodiments of the present application, the data processing flow during the training of the constructed virtualized statistical channel model is the same as the data processing flow for actual channel estimation using the trained virtualized statistical channel model, therefore, in this embodiment, in order not to perform repetitive redundant descriptions, only the data processing flow for actual channel estimation is used as an example for detailed description below.

[0084] Step 102: Input the cell location and cell antenna parameters into the cell feature encoder for feature encoding processing to obtain the cell feature encoding.

[0085] Step 103: Input the grid location and cell location into the grid feature encoder for feature encoding processing to obtain the grid feature encoding.

[0086] The following gives a detailed description of Steps 102 to 103.

[0087] In the realistic statistical channel model as Figure 3 shown, in order to extract features of different environmental information, encoder designs are carried out from two dimensions of the grid and the communication cell respectively.

[0088] Among them, the grid feature encoder is used to extract the geographical information of the grid in the propagation environment. Considering that the propagation environment satisfies translational and rotational invariance in the geographical space, the relative coordinates of the grid are calculated according to the absolute coordinates of the grid and the communication cell, and used as the location feature, as shown in the following formula (12).

[0089] (12) Where and are the absolute coordinates of the th grid and the th communication cell.

[0090] In wireless signal propagation, the influence of the propagation distance on signal attenuation is the most direct and obvious. Further, calculate the propagation distance and relative angle between the grid and the cell, and use them to enhance the representation ability of the physical features. The mechanical horizontal angle, mechanical downward tilt angle and frequency point between the th grid and the th communication cell are respectively as shown in the following formulas (13), (14) and (15).

[0091] (13) (14) (15) Finally, in the grid feature encoder, the above features are concatenated into the input features of the grid feature encoder, as shown in the following formula (16).

[0092] (16) Where, as can be seen from formula (12), is the difference in the x-axis coordinates between the th grid and the th communication cell, is the The y-axis coordinate difference between the th grid and the th communication cell, and the z-axis coordinate difference between the th grid and the

[0093] th communication cell. In addition, the cell feature encoder is used to extract the common parameter information of the cell. As can be seen from formula (4), the propagation of the signal is jointly determined by the coefficient matrix and the channel angular power spectrum

[0094] (17) where the propagation environment affects the channel angular power spectrum, and the antenna parameters of the cell affect the coefficient matrix. Therefore, the position and antenna parameters of the cell are used as the input features of the cell feature encoder, as shown in formula (17) below. In some embodiments, in practical applications, after obtaining the grid position of the grid to be estimated, the cell position of the communication cell, and the cell antenna parameters, the input features as shown in formula (16) above are respectively used to input the grid encoder for feature extraction to obtain the grid feature encoding and the input features as shown in formula (17) are used to input the cell encoder for feature extraction to obtain the cell feature encoding

[0095] Refer to Figure 5 , which is a schematic structural diagram of a cell feature encoder and a grid feature encoder provided by an embodiment of the present application. As Figure 5 shown, it shows the structure and process of the feature encoder in the realistic statistical channel model. Specifically, it includes two parallel encoding processing paths: on the left is the grid feature encoder, which receives the grid feature input related to a specific communication cell q (which may include information such as relative coordinates, distance, angle, etc.), and sequentially passes through a fully connected layer, a rectified linear unit (ReLU) activation function, a layer normalization layer, a second fully connected layer, and a second layer normalization layer for deep feature extraction; on the right is the cell feature encoder, which receives the feature input of the communication cell q itself (such as cell position, antenna parameters, etc.), and sequentially passes through a fully connected layer, a rectified linear unit (ReLU), and a layer normalization layer for feature encoding. Finally, the grid feature encoding and the cell feature encoding output by the two encoders are concatenated at the operation point marked as "C" (Concatenation) to merge two different but related feature information into a richer feature vector as the input for subsequent processing.

[0096] Step 104: Input the cell feature encoding and the grid feature encoding into the angular power spectrum estimation network for power spectrum estimation to obtain the estimated channel power spectrum between the grid to be estimated and the communication cell.

[0097] The following provides a detailed description of Step 104.

[0098] In some embodiments, after obtaining the grid feature encoding and the cell feature encoding , the grid feature encoding and the cell feature encoding will be further used, along with the angular power spectrum estimation network, for power spectrum estimation to obtain the real-time estimated channel power spectrum between the grid to be estimated and the communication cell , as described in detail below.

[0099] Referring to Figure 6 , inputting the cell feature encoding and the grid feature encoding into the angular power spectrum estimation network for power spectrum estimation to obtain the estimated channel power spectrum between the grid to be estimated and the communication cell includes the following steps 601 to 603.

[0100] Step 601: After performing an exclusive OR operation on the cell feature encoding and the grid feature encoding, superimpose the position embedding of the grid position to obtain a feature processing unit.

[0101] The following provides a detailed description of Step 601.

[0102] In some embodiments, Transformer uses global information for parallel processing, enabling the virtualized statistical channel model to efficiently process long sequence inputs. However, in natural language processing, the position of different tokens in the sequence is crucial. Therefore, position encoding is usually adopted to identify the absolute or relative position of tokens in the sequence.

[0103] In the channel estimation task corresponding to the embodiments of the present application, multi-grid sequence modeling is adopted, and each Token corresponds to a grid-cell pair. To characterize the spatial association between different grids, a frequency-based sine-cosine geospatial position embedding method is introduced. This method does not increase additional training parameters and can enhance the model's understanding of spatial relationships.

[0104] Therefore, the position embedding of the grid position is as shown in the following formula (18).

[0105] (18) where , that is correspond to the position embeddings of the x-axis, y-axis, and z-axis respectively. is the dimension of the input token.

[0106] Therefore, after performing an exclusive OR operation on the cell feature encoding and the grid feature encoding , and then superimposing the position embedding of the grid position , the feature processing unit is obtained as shown in the following formula (19).

[0107] (19) Step 602: Input the feature processing unit into the transformer module for attention processing to obtain the estimated processed data.

[0108] The following is a detailed description of step 602.

[0109] Next, in order for the model to capture the associations between each grid in the multi-grid sequence (the consistency of the wireless channel in the geographical space), regardless of whether these relationships are local or global, the self-attention mechanism is used to enable the model to dynamically focus on the information at other positions in the sequence when inputting each grid of the multi-grid sequence, thereby effectively jointly performing structured feature extraction on the multi-grid. Therefore, the feature processing unit is further input into the transformer module for attention processing to obtain the estimated processed data , which is described in detail as follows.

[0110] Referring to Figure 7 , inputting the feature processing unit into the transformer module for attention processing to obtain the estimated processed data includes the following steps 701 to 702.

[0111] Step 701: Obtain other feature processing units between other grids in the target area and the communication cell, and combine the other feature processing units and the feature processing unit to obtain a feature unit matrix.

[0112] Step 702: Input the feature unit matrix into the multi-head attention module for multi-head attention processing to obtain the estimated processed data.

[0113] The following is a detailed description of steps 701 to 702.

[0114] In some embodiments, in the transformer module, first based on other feature processing units between other grids in the target area and the communication cell, and combine the other feature processing units and the feature processing unit, the obtained feature unit matrix is input as the token matrix of the multi-grid sequence, that is, the feature unit matrix ( ), by using the linear transformation matrix ( ), ( ) and ( ), successively obtaining matrix (Query), (Key) and (Value) for facilitating attention processing.

[0115] In addition, multi - head self - attention is an extension of self - attention. By introducing multiple "heads", the model can perform self - attention calculations in parallel in different sub - spaces while keeping the total number of parameters unchanged, thereby capturing more diverse features and relationships and suppressing overfitting to some extent. Therefore, a multi - head attention module is set in the transformer module, and multiple self - attention processing modules are set in the multi - head attention module.

[0116] Referring to Figure 8 , it is a schematic structural diagram of a multi - head attention module provided by an embodiment of the present application. As shown in Figure 8 , it details the structure and process in the multi - head attention module, which is the core component of the transformer module. Its process is as follows: First, the input query (Q), key (K), and value (V) vectors (these vectors are usually representations of the previous layer or input embeddings) are respectively subjected to h independent linear transformations (the "linear transformation" layer at the bottom of the figure, with h groups of parallel transformations) to generate respective Q, K, V representations for h different self - attention processing modules; then, inside each self - attention processing module, a self - attention calculation (usually scaled dot - product attention) is performed using its exclusive Q, K, V, enabling each position to attend to the information of all positions in the input sequence and calculate a weighted representation; subsequently, the output vectors obtained by parallel calculation of the h self - attention processing modules are concatenated to combine the different sub - space information or dependencies captured by different attention heads; finally, the concatenated vector passes through a final linear transformation layer for integration and dimensional adjustment to obtain the final output Z of the multi - head self - attention module. This output integrates information from multiple different representation sub - spaces and can more comprehensively capture the spatial channel feature dependencies between different grid positions in the input sequence.

[0117] Therefore, after obtaining the feature unit matrix ( ), it is input into the multi - head attention module for multi - head attention processing to obtain the estimated processed data , which is described in detail as follows.

[0118] Referring to Figure 9, input the feature unit matrix into the multi-head attention module for multi-head attention processing to obtain the estimated processed data, including the following steps 901 to 904.

[0119] Step 901: In the multi-head attention module, based on the ratio of the data dimension of the feature unit matrix to the number of attention heads, obtain the division dimension value of each attention head.

[0120] Step 902: Generate the projection parameter matrix of each attention head based on the division dimension value.

[0121] Step 903: In each self-attention processing module, generate a matrix query vector, a matrix key vector, and a matrix value vector based on the feature unit matrix, and perform attention processing on the matrix query vector, the matrix key vector, and the matrix value vector based on the corresponding projection parameter matrix and linear transformation matrix to obtain sub-estimated processed data.

[0122] Step 904: Perform multi-head attention processing on multiple sub-estimated processed data to obtain the estimated processed data.

[0123] The following is a detailed description of steps 901 to 904.

[0124] In some embodiments, in the multi-head attention module as Figure 8 shown, first, based on the data dimension of the feature unit matrix and the number of attention heads ratio, obtain the division dimension value of each attention head , and then, generate the projection parameter matrix of each attention head based on this division dimension value, including , , and , is the number of self-attention processing modules in the multi-head attention module. Based on this, the multi-head attention processing flow in the multi-head attention module is as shown in the following formula (20).

[0125] (20) Next, in the multi-head attention module, perform attention processing on the feature unit matrix based on multiple self-attention processing modules together. Refer to Figure 10 , which is a schematic structural diagram of a self-attention processing module provided by an embodiment of the present application. As Figure 10As shown, the internal structure and data processing flow of the Scaled Dot-Product Attention mechanism are detailed. This is the core computational unit that constitutes each "head" in the Multi-Head Attention. This process receives three inputs: the query matrix (Q), the key matrix (K), and the value matrix (V). First, the query matrix Q and the key matrix K are "matrix multiplied" (usually calculating the dot product of Q and the transpose of K) to obtain the raw attention scores, which measure the similarity or correlation between each query and each key. Then, these raw scores go through a "scaling" step, usually dividing by the square root of the key vector dimension to prevent the gradient from being too small or too large and to stabilize the training process. The scaled attention scores are then fed into a "normalization" layer, usually applying the Softmax function to transform the scores into a probability distribution (i.e., attention weights), such that the sum of the weights for all keys corresponding to each query is 1. Finally, these normalized attention weights are "matrix multiplied" with the value matrix V to calculate the weighted sum of the values. This final output Z is the result of aggregating the value information weighted by the similarity between the queries and the keys, representing the context-aware feature representation after being processed by the attention mechanism.

[0126] As Figure 10 shown, in each self-attention processing module, a matrix query vector, a matrix key vector, and a matrix value vector are generated based on the feature unit matrix, and attention processing is performed on the matrix query vector, the matrix key vector, and the matrix value vector based on the corresponding projection parameter matrix and linear transformation matrix to obtain the sub-estimation processing data. That is, the input is the feature unit matrix of the multi-grid sequence ( ), and by using the linear transformation matrix ( ), ( ), and ( ), the matrix (Query), (Key), and (Value) are obtained in sequence. , , The output value, that is, the sub-estimation processing data, is obtained through the following formula (21) .

[0127] (21) where is a scaling factor, which is used for normalization. This method of determining the weight distribution of and by their similarity is called dot - product attention. Self - attention adaptively calculates weights in parallel and can effectively capture the correlations between any two grids in the sequence. After that, based on the multi - head attention processing flow as shown in the above formula (20), multi - head attention is performed on the processed data of multiple sub - estimates to obtain the processed data of the estimate.

[0128] Through the above steps 701 to 702, and steps 901 to 904, by combining the feature units of multiple grids into a feature unit matrix and then applying the multi - head self - attention mechanism, the model can make the feature representation (as query Q) of each grid focus on and weighted - aggregate the information from all other grids (including itself), effectively learning complex spatial patterns such as how signals propagate in space, the spatial continuity of shadow effects, and the mutual influence of interference between different positions. The multi - head mechanism further allows the model to capture multiple different types or scales of spatial correlation features simultaneously from different representation sub - spaces. This deep modeling ability of spatial context overcomes the limitations of traditional methods or simple models that may regard each grid as an isolated point or only consider local neighborhood effects, enabling the finally output processed data of the estimate to more comprehensively and realistically reflect the spatial heterogeneity and interactions of the large - scale wireless environment, greatly improving the accuracy and "real - scene" degree of subsequent channel power spectrum estimation, and providing higher - quality input for applications such as network optimization that rely on accurate channel information.

[0129] Step 603: Input the processed data of the estimate into the angular power spectrum estimation network for power spectrum estimation to obtain the estimated channel power spectrum.

[0130] The following provides a detailed description of step 603.

[0131] In some embodiments, in the processing flow of the real - scene statistical channel model as shown in

[0132] , after obtaining the processed data of the estimate output by the multi - head attention module, the processed data of the estimate is further input into the angular power spectrum estimation network for power spectrum estimation to obtain the estimated channel power spectrum, which is described in detail as follows. Figure 3 Referring to

[0133] , inputting the processed data of the estimate into the angular power spectrum estimation network for power spectrum estimation to obtain the estimated channel power spectrum includes the following steps 1101 to 1104. Figure 11

[0134] ​Step 1101: In the angular power spectrum estimation network, sequentially perform linear processing, activation function processing, layer normalization processing, and regularization processing on the estimated processing data to obtain preliminarily processed data.

[0135] Step 1102: Sequentially perform residual connection enhancement processing and feature transfer processing on the preliminary processing result to obtain enhanced processing data.

[0136] Step 1103: Perform non-negative activation processing on the enhanced processing data to obtain activated processing data.

[0137] Step 1104: Based on the activated processing data, perform estimation processing to obtain the estimated channel power spectrum.

[0138] The following provides a detailed description of Steps 1101 to 1104.

[0139] In some embodiments, considering the characteristics of non-negative sparse high-dimensional angular power spectrum, an angular power spectrum estimation network is designed. Refer to Figure 12 , which is a schematic structural diagram of an angular power spectrum estimation network provided by an embodiment of the present application. As Figure 12 shown, progressive dimension expansion is adopted. The basic components of each layer include a linear layer, a GELU activation function, layer normalization, and Dropout. Residual connections are used in the second and third layers to enhance gradient flow and feature transfer. Finally, a ReLU activation function is used in the output layer to ensure non-negative output while promoting sparsity.

[0140] Therefore, based on the angular power spectrum estimation network as Figure 12 shown, first, sequentially perform linear processing, activation function processing, layer normalization processing, and regularization processing on the estimated processing data to obtain preliminarily processed data, so as to introduce non-linearity while accelerating training and improving the generalization ability of the model, and thus reducing the dependence relationship between neurons; then, sequentially perform residual connection enhancement processing and feature transfer processing on the preliminary processing result to obtain enhanced processing data, enabling the model to alleviate the problem of gradient disappearance, enabling the model to more easily learn the identity mapping, and enabling the model to utilize earlier feature information; next, perform non-negative activation processing on the enhanced processing data to obtain activated processing data, enabling the model to ensure the rationality of the output; finally, based on the activated processing data, perform estimation processing to obtain the estimated channel power spectrum, and the dimension of the output depends on the representation method of the channel power spectrum. For example, if the channel power spectrum is represented as a vector, the dimension of the output is the length of the vector.

[0141] Through the above steps 1101 to 1104, the input features are preliminarily subjected to deep non-linear transformation and refinement through linear transformation, non-linear activation (such as GELU), layer normalization, and regularization (such as Dropout), while ensuring the stability of training and the generalization ability of the model. Residual connections and further feature transfer processing are adopted, allowing the construction of deeper network structures to capture finer feature details, effectively solving the problem of vanishing gradients in the training of deep networks. Thus, the ability of the model to learn and infer precise power spectrum details from complex input features is significantly improved. Additionally, the result before the final output is processed using a non-negative activation function (such as ReLU), which forcibly ensures that the estimated power spectrum values are all non-negative, directly reflecting the constraints of the physical model, making the output results more in line with physical reality, and improving the reliability and credibility of the estimation results. Based on this activation data that has undergone deep processing and physical constraints, the final estimated channel power spectrum is directly obtained, providing an end-to-end, precise mapping from high-level features to specific power spectrum distributions, and providing a high-quality, physically meaningful description of channel characteristics for subsequent network optimization applications.

[0142] Refer to Figure 13 , which is a schematic diagram of the concept of a channel power spectrum estimation provided by an embodiment of the present application. As Figure 13 shown, by implementing the channel power spectrum estimation method provided by the present application, at the input end, the realistic statistical channel model receives two types of information: The first type is parameters, including key information describing the physical scene configuration, such as the geographical coordinates of multiple grids in the target area (multi-grid coordinates), the geographical coordinates of multiple communication cells participating in communication, and the specific antenna setting parameters of these cells, such as antenna gain, downtilt angle, azimuth angle (simplified as "and antenna gain angle" in the figure), and operating frequency points, etc.; The second type is data, specifically referring to the reference signal received power (RSRP) measured in the actual environment. These data reflect the impact of the real wireless propagation environment on the signal. The "+" sign in the figure indicates that these two types of inputs are combined and utilized during model training or use. These input information are sent to the core "large-scale realistic statistical channel model" for processing. This model aims to learn the complex mapping relationship between parameter configurations, real RSRP data, and underlying channel statistical characteristics.

[0143] At the output end, the virtualized statistical channel model mainly generates "channel angular power spectrum statistical characteristics of multi-cell and multi-grid". This means that the model can predict the angular power spectrum (APS) of the channel between any specified grid and the relevant cells within the target area, that is, the distribution of signal energy at different angles of arrival, which is a key channel statistical characteristic. In addition, the figure also mentions that the output contains information related to "large-scale environmental multipath topology", which means that the virtualized statistical channel model can also implicitly or explicitly learn and reflect the characteristics of the multipath propagation structure in a large-scale environment (that is, after obtaining the channel power spectrum, combined with formula (6) to predict the maximum value of the received power of the reference signal after adjusting the network parameters in the wireless communication system, so as to accurately estimate the channel of the wireless communication system), or rather, the accurate APS output itself contains an understanding of the multipath topology.

[0144] To improve the reliability of a channel power spectrum estimation method provided in an embodiment of the present application, simulation verification is performed on the channel power spectrum estimation. The measured data set covers 10,012 grids of 10m x 10m in 380 communication cells in a certain area, and a total of eight rounds of data are included. The codebooks of several cells are switched in each round of data. We use the 1st, 3rd, 5th, and 7th rounds as the training set, and the 2nd, 4th, 6th, and 8th rounds as the test set. The multi-grid sequence length is set to 20, and the number of model parameters is 6,975,736 (6.98M). We measure the model prediction performance by the mean absolute error.

[0145] Refer to Figure 14 , which is a schematic diagram of the loss decline curve simulation of a virtualized statistical channel model provided in an embodiment of the present application. As Figure 14 shown in, it is the loss decline curve of the model trained for 50 epochs (about 5 minutes), where the yellow represents the training loss. It can be seen that the training loss of the virtualized statistical channel model provided in the present application declines rapidly and smoothly, and at the same time, the test loss also declines accordingly, which proves the effectiveness of our model design.

[0146] In addition, in this embodiment, the channel power spectrum estimation method proposed in the present application is compared with other baseline algorithms, including the iterative threshold soft threshold algorithm (ISTA), the non-negative orthogonal matching pursuit algorithm (NNOMP), and the weighted non-negative orthogonal matching pursuit algorithm (WNNOMP).

[0147] Refer to Figure 15 , which is a schematic diagram of the performance simulation of a channel power spectrum estimation method provided in an embodiment of the present application. As Figure 15As shown, the channel power spectrum estimation method proposed in this application not only ensures training efficiency but also outperforms all baseline algorithms in terms of performance. It should be noted that the realistic statistical channel model proposed in this application can predict grids not included in the training set, while other optimization algorithms can only make predictions based on existing grids, and the generalization ability of the realistic statistical channel model proposed in this application is further improved.

[0148] The channel power spectrum estimation method, device, electronic device and storage medium proposed in the embodiments of the present application are executed by a virtualized statistical channel model. The virtualized statistical channel model at least includes: a cell feature encoder, a grid feature encoder, and an angular power spectrum estimation network. The method includes: First, obtain the grid position of the grid to be estimated in the target area, and obtain the cell position and cell antenna parameters of at least one communication cell corresponding to the grid position. The virtualized statistical channel model is pre-trained by the reference signal received power between multiple grids and communication cells in the target area. Among them, the training steps of the virtualized statistical channel model include: obtaining the channel impulse response function and precoding matrix between each antenna in the communication cell and each grid, accumulating the product of all channel impulse response functions and precoding matrices, and performing a square process, then multiplying by the transmit power of the communication cell to obtain the reference signal received power function of each grid, performing a time expectation process on the reference signal received power function to obtain the expected reference signal received power function. The expected reference signal received power function includes the coefficient matrix between the grid and the communication cell. Based on the coefficient matrix, the reference signal received power function, and the channel power spectrum estimation function between each grid and the communication cell, generate a channel power spectrum estimation model. Based on the channel power spectrum estimation model and multiple reference signal received powers, obtain the sparse penalty weight for the virtualized statistical channel model. Based on the product of the sparse penalty weight and the channel power spectrum estimation function, obtain the penalty channel power spectrum function. Based on the accumulated value of the channel power spectrum estimation model and the penalty channel power spectrum function, generate a loss function. Based on the loss function and multiple reference signal received powers, perform multiple iterative updates on the virtualized statistical channel model; Then, input the cell position and cell antenna parameters into the cell feature encoder for feature encoding processing to obtain the cell feature encoding; Next, input the grid position and cell position into the grid feature encoder for feature encoding processing to obtain the grid feature encoding;Finally, after performing an exclusive OR operation on the cell feature encoding and the grid feature encoding, and then overlaying the position embedding of the grid position, a feature processing unit is obtained. Other feature processing units between other grids in the target area and the communication cell are acquired, and the other feature processing units and the feature processing unit are combined to obtain a feature unit matrix. In the multi-head attention module, based on the ratio of the data dimension of the feature unit matrix to the number of attention heads, the division dimension value of each attention head is obtained, and a projection parameter matrix for each attention head is generated based on the division dimension value. In each self-attention processing module, a matrix query vector, a matrix key vector, and a matrix value vector are generated based on the feature unit matrix, and attention processing is performed on the matrix query vector, the matrix key vector, and the matrix value vector based on the corresponding projection parameter matrix and linear transformation matrix to obtain sub-estimation processing data. Multi-head attention processing is performed on multiple sub-estimation processing data to obtain estimation processing data. In the angular power spectrum estimation network, the estimation processing data is sequentially subjected to linear processing, activation function processing, layer normalization processing, and regularization processing to obtain preliminary processing data. The preliminary processing result is sequentially subjected to residual connection enhancement processing and feature transfer processing to obtain enhanced processing data. The enhanced processing data is subjected to non-negative activation processing to obtain activation processing data. Estimation processing is performed based on the activation processing data to obtain an estimated channel power spectrum.

[0149] In the embodiments of the present application, a real-scene statistical channel model is trained by using the measured reference signal received power data with low acquisition cost between the target area and the communication cell in advance and adopting a method that combines physical model guidance and deep learning, avoiding the real-time, complex, and high-dimensional environmental information acquisition and high-computation deterministic modeling required by traditional high-precision methods, so as to reduce the training complexity of the estimation model for channel power spectrum estimation and improve the training efficiency. And because the model directly learns from the measured reference signal received power data that reflects the influence of the real environment, it is possible to generate a channel power spectrum estimation result that is closer to the actual situation and has more "real-scene" characteristics. In actual channel power spectrum estimation, efficient feature extraction is carried out by using the grid position of the grid to be estimated that can be easily obtained in real time, the cell position and cell antenna parameters in the communication cell, and the pre-obtained real-scene statistical channel model, thus avoiding the need for real-time acquisition of complex and time-varying high-precision environmental information, significantly reducing the computational complexity, and efficiently obtaining the real-time estimated channel power spectrum in the wireless communication system. Furthermore, while ensuring the accuracy of channel estimation, the channel estimation efficiency in the wireless communication system is effectively improved. In addition, starting from basic physical quantities such as channel impulse response, precoding, and transmit power, the expected reference signal received power is deduced to have a clear mathematical relationship with the underlying channel statistical characteristics (ultimately reflected as the angle power spectrum and the coefficient matrix), which provides a solid physical basis and interpretable basis for subsequent model training, avoiding the drawback of the pure black-box model lacking physical meaning. And by using this physical relationship to construct the learning objective and iteratively training the deep learning model with the actually measured and easily obtained reference signal received power data, the model can learn the complex propagation characteristics in a specific "real-scene" environment, overcoming the problems of poor generalization ability of traditional statistical models or large computational complexity of deterministic models that rely on high-precision environmental information, realizing the "real-scene" and efficiency of the model. And by introducing a sparse penalty term into the loss function, the prior knowledge that the angle power spectrum of the wireless channel usually has sparsity is utilized to guide the model to learn a more physically realistic, simpler, and more effective power spectrum representation, which not only helps to improve the generalization ability and estimation accuracy of the model, but also can improve the robustness of the model to a certain extent, thus obtaining an efficient, accurate, and practical channel power spectrum estimation model that combines physical principles, measured data, and prior knowledge. And by combining the feature units of multiple grids into a feature unit matrix and then applying the multi-head self-attention mechanism, the model can make the feature representation of each grid (as the query Q) focus on and weighted aggregate the information from all other grids (including itself), effectively learning complex spatial patterns such as how signals propagate in space, the spatial continuity of shadow effects, and the mutual influence of interference between different positions. The multi-head mechanism further allows the model to simultaneously capture multiple different types or scales of spatial correlation features from different representation subspaces.This deep modeling ability of spatial context overcomes the limitations of traditional methods or simple models that may treat each grid as an isolated point or only consider the influence of local neighbors, enabling the estimated processing data of the final output to more comprehensively and realistically reflect the spatial heterogeneity and interactions of the large-scale wireless environment. This greatly improves the accuracy and "real-scene" degree of subsequent channel power spectrum estimation, providing higher-quality input for applications such as network optimization that rely on accurate channel information. Moreover, through linear transformation, non-linear activation (such as GELU), layer normalization, and regularization (such as Dropout), the input features are initially subjected to deep non-linear transformation and refinement, while ensuring the stability of training and the generalization ability of the model. By using residual connections and further feature transfer processing, it allows for the construction of deeper network structures to capture finer feature details, effectively solving the vanishing gradient problem in deep network training. Consequently, it significantly enhances the model's ability to learn and infer precise power spectrum details from complex input features. Additionally, processing the results before the final output with a non-negative activation function (such as ReLU) forcibly ensures that the estimated power spectrum values are all non-negative, which directly reflects the constraints of the physical model, making the output results more in line with physical reality and improving the reliability and credibility of the estimation results. Based on this activation data that has been deeply processed and physically constrained, the final estimated channel power spectrum is directly obtained, providing an end-to-end, precise mapping from high-level features to specific power spectrum distributions, and providing a high-quality, physically meaningful channel characteristic description for subsequent network optimization applications.

[0150] The embodiment of the present application also provides a channel power spectrum estimation device, which can implement the above channel power spectrum estimation method. Referring to Figure 16 , the device 1600 includes: A data acquisition module 1610, configured to acquire the grid position of the grid to be estimated in the target area, and acquire the cell positions and cell antenna parameters of at least one communication cell corresponding to the grid position. The real-scene statistical channel model is pre-trained by the reference signal received power between multiple grids and communication cells in the target area; A cell feature processing module 1620, configured to input the cell position and cell antenna parameters into a cell feature encoder for feature encoding processing to obtain cell feature encoding; A grid feature processing module 1630, configured to input the grid position and cell position into a grid feature encoder for feature encoding processing to obtain grid feature encoding; A channel power spectrum estimation module 1640, configured to input the cell feature encoding and grid feature encoding into an angular power spectrum estimation network for power spectrum estimation to obtain the estimated channel power spectrum between the grid to be estimated and the communication cell.

[0151] In some embodiments, the channel power spectrum estimation module 1640 is further configured to: After performing exclusive OR processing on the cell feature encoding and the grid feature encoding, and then superimposing the position embedding of the grid position, a feature processing unit is obtained; Input the feature processing unit into a transformer module for attention processing to obtain estimated processing data; Input the estimated processing data into an angular power spectrum estimation network for power spectrum estimation to obtain an estimated channel power spectrum.

[0152] In some embodiments, the channel power spectrum estimation module 1640 is further configured to: Obtain other feature processing units between other grids and communication cells in the target area, and combine the other feature processing units and the feature processing unit to obtain a feature unit matrix; Input the feature unit matrix into a multi-head attention module for multi-head attention processing to obtain estimated processing data.

[0153] In some embodiments, the channel power spectrum estimation module 1640 is further configured to: In the multi-head attention module, based on the ratio of the data dimension of the feature unit matrix to the number of attention heads, obtain the division dimension value of each attention head; Generate a projection parameter matrix for each attention head based on the division dimension value; In each self-attention processing module, generate a matrix query vector, a matrix key vector, and a matrix value vector based on the feature unit matrix, and perform attention processing on the matrix query vector, the matrix key vector, and the matrix value vector based on the corresponding projection parameter matrix and linear transformation matrix to obtain sub-estimated processing data; Perform multi-head attention processing on multiple sub-estimated processing data to obtain estimated processing data.

[0154] In some embodiments, the channel power spectrum estimation module 1640 is further configured to: In the angular power spectrum estimation network, perform linear processing, activation function processing, layer normalization processing, and regularization processing on the estimated processing data in sequence to obtain preliminary processing data; Perform residual connection enhancement processing and feature transfer processing on the preliminary processing result in sequence to obtain enhanced processing data; Perform non-negative activation processing on the enhanced processing data to obtain activated processing data; Perform estimation processing based on the activated processing data to obtain an estimated channel power spectrum.

[0155] In some embodiments, the data acquisition module 1610 is further configured to: Obtain the channel impulse response function and the precoding matrix between each antenna and each grid in the communication cell; Sum the products of all channel impulse response functions and the precoding matrix, square the result, and then multiply by the transmit power of the communication cell to obtain the reference signal received power function for each grid; Perform time expectation processing on the reference signal received power function to obtain the expected reference signal received power function, which includes the coefficient matrix between the grid and the communication cell; Generate a channel power spectrum estimation model based on the coefficient matrix, the reference signal received power function, and the channel power spectrum estimation function between each grid and the communication cell; Iteratively update the realistic statistical channel model multiple times based on the channel power spectrum estimation model and multiple reference signal received powers.

[0156] In some embodiments, the data acquisition module 1610 is further configured to: Obtain a sparse penalty weight, and based on the product of the sparse penalty weight and the channel power spectrum estimation function, obtain a penalty channel power spectrum function; Generate a loss function based on the accumulated value of the channel power spectrum estimation model and the penalty channel power spectrum function; Iteratively update the realistic statistical channel model multiple times based on the loss function and multiple reference signal received powers.

[0157] In the above embodiments, the descriptions of the various embodiments have their own focuses. For the parts not detailed in a certain embodiment, the specific implementation of the channel power spectrum estimation device is basically the same as the specific implementation of the above channel power spectrum estimation method, and will not be elaborated here.

[0158] In the embodiments of the present application, the channel power spectrum estimation device uses the measured reference signal received power data with low acquisition cost between the target area and the communication cell in advance, and adopts a method that combines physical model guidance and deep learning for training to obtain a real-scene statistical channel model, avoiding the real-time, complex, high-dimensional environmental information acquisition and high-computation deterministic modeling required by traditional high-precision methods, so as to reduce the training complexity of the estimation model for channel power spectrum estimation and improve the training efficiency. Since the model directly learns from the measured reference signal received power data reflecting the real environmental impact, it enables the generation of a channel power spectrum estimation result that is closer to the actual situation and has more "real-scene" characteristics. In actual channel power spectrum estimation, efficient feature extraction is carried out by using the grid position of the grid to be estimated that can be easily obtained in real time, the cell position and cell antenna parameters in the communication cell, and the previously obtained real-scene statistical channel model, thus avoiding the need for real-time acquisition of complex and time-varying high-precision environmental information, significantly reducing the computational complexity, and efficiently obtaining the real-time estimated channel power spectrum in the wireless communication system. Furthermore, while ensuring the accuracy of channel estimation, the channel estimation efficiency in the wireless communication system is effectively improved. In addition, starting from basic physical quantities such as channel impulse response, precoding, and transmit power, the explicit mathematical relationship between the expected reference signal received power and the underlying channel statistical characteristics (ultimately reflected as the angle power spectrum and the coefficient matrix) is derived, which provides a solid physical basis and interpretable basis for subsequent model training, avoiding the drawback of the pure black box model lacking physical meaning. Using this physical relationship to construct the learning objective and iteratively training the deep learning model with the actually measured and easily obtained reference signal received power data enables the model to learn the complex propagation characteristics in a specific "real-scene" environment, overcoming the problems of poor generalization ability of traditional statistical models or large computational complexity of deterministic models relying on high-precision environmental information, realizing the "real-scene" and efficiency of the model, and by introducing a sparse penalty term into the loss function, using the prior knowledge that the angle power spectrum of the wireless channel usually has sparsity, guiding the model to learn a more physically realistic, simpler and more effective power spectrum representation, which not only helps to improve the generalization ability and estimation accuracy of the model, but also can improve the robustness of the model to a certain extent, thus obtaining an efficient, accurate and practical channel power spectrum estimation model that combines physical principles, measured data and prior knowledge; and by combining the feature units of multiple grids into a feature unit matrix and then applying the multi-head self-attention mechanism, the model can make the feature representation of each grid (as the query Q) focus on and weighted aggregate the information from all other grids (including itself), effectively learning complex spatial patterns such as how signals propagate in space, the spatial continuity of shadow effects, and the mutual influence of interference between different positions. The multi-head mechanism further allows the model to simultaneously capture multiple different types or scales of spatial correlation features from different representation subspaces.This deep modeling ability of spatial context overcomes the limitations of traditional methods or simple models that may regard each grid as an isolated point or only consider the influence of local neighbors, enabling the finally output estimated processing data to more comprehensively and realistically reflect the spatial heterogeneity and interactions of large-scale wireless environments, greatly enhancing the accuracy and "realistic" degree of subsequent channel power spectrum estimation, and providing higher-quality input for applications such as network optimization that rely on accurate channel information. Moreover, through linear transformation, non-linear activation (such as GELU), layer normalization, and regularization (such as Dropout), the input features are initially subjected to deep non-linear transformation and refinement, while ensuring the stability of training and the generalization ability of the model. Residual connections and further feature passing processing are adopted, allowing the construction of deeper network structures to capture finer feature details, effectively solving the problem of gradient disappearance in deep network training, thereby significantly enhancing the model's ability to learn and infer precise power spectrum details from complex input features. Additionally, applying a non-negative activation function (such as ReLU) to process the results before the final output forcibly ensures that the estimated power spectrum values are all non-negative, which directly reflects the constraints of the physical model, making the output results more in line with physical reality, improving the reliability and credibility of the estimation results. Based on this deeply processed and physically constrained activation data, the final estimated channel power spectrum is directly obtained, providing an end-to-end, precise mapping from high-level features to specific power spectrum distributions, and providing a high-quality, physically meaningful description of channel characteristics for subsequent network optimization applications.

[0159] The embodiments of the present application further provide an electronic device, including: At least one memory; At least one processor; At least one program; The program is stored in the memory, and the processor executes the at least one program to implement the channel power spectrum estimation method described above in the present application. This electronic device can be any intelligent terminal including a mobile phone, a tablet computer, a personal digital assistant (Personal Digital Assistant, abbreviated as PDA), an in-vehicle computer, etc.

[0160] Please refer to Figure 17 , Figure 17 which schematically shows the hardware structure of an electronic device in another embodiment. The electronic device includes: The processor 1701 can be implemented in the form of a general-purpose CPU (Central Processing Unit), a microprocessor, an application-specific integrated circuit (ASIC), or one or more integrated circuits, etc., and is used to execute relevant programs to implement the technical solutions provided in the embodiments of the present application; The memory 1702 can be implemented in the form of a ROM (ReadOnly Memory), a static storage device, a dynamic storage device, or a RAM (Random Access Memory), etc. The memory 1702 can store an operating system and other application programs. When implementing the technical solutions provided in the embodiments of this specification through software or firmware, the relevant program codes are stored in the memory 1702 and are called by the processor 1701 to execute the channel power spectrum estimation method in the embodiments of the present application; The input / output interface 1703 is used to implement information input and output; The communication interface 1704 is used to implement communication interaction between this device and other devices, and can implement communication through wired means (such as USB, network cable, etc.) or through wireless means (such as mobile network, WIFI, Bluetooth, etc.); The bus 1705 transmits information between the various components of the device (such as the processor 1701, the memory 1702, the input / output interface 1703, and the communication interface 1704); Among them, the processor 1701, the memory 1702, the input / output interface 1703, and the communication interface 1704 are communicatively connected to each other inside the device through the bus 1705.

[0161] The embodiments of the present application also provide a storage medium, which is a computer-readable storage medium. The storage medium stores a computer program, and when the computer program is executed by a processor, the above-mentioned channel power spectrum estimation method is implemented.

[0162] As a non-transitory computer-readable storage medium, the memory can be used to store non-transitory software programs and non-transitory computer-executable programs. In addition, the memory can include high-speed random access memory, and can also include non-transitory memory, such as at least one magnetic disk storage device, a flash memory device, or other non-transitory solid-state storage devices. In some embodiments, the memory may optionally include a memory remotely set relative to the processor, and these remote memories can be connected to the processor through a network. Examples of the above networks include, but are not limited to, the Internet, an enterprise intranet, a local area network, a mobile communication network, and combinations thereof.

[0163] The embodiments described in the embodiments of the present application are for more clearly illustrating the technical solutions of the embodiments of the present application, and do not constitute a limitation on the technical solutions provided by the embodiments of the present application. Those skilled in the art can know that with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of the present application are equally applicable to similar technical problems.

[0164] Those skilled in the art can understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of the present application, and may include more or fewer steps than shown in the figures, or combine certain steps, or different steps.

[0165] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, that is, they may be located in one place, or may be distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0166] Those of ordinary skill in the art can understand that all or some of the steps in the methods disclosed above, and the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, and their appropriate combinations.

[0167] The terms "first", "second", "third", "fourth", etc. (if any) in the specification of the present application and the above-mentioned drawings are used to distinguish similar objects, and do not have to be used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments of the present application described here can be implemented in an order other than those illustrated or described here. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units does not have to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products, or devices.

[0168] It should be understood that in this application, "at least one (item)" means one or more, and "a plurality" means two or more. "And / or" is used to describe the association relationship of associated objects, indicating that three relationships can exist. For example, "A and / or B" can mean: only A exists, only B exists, and both A and B exist at the same time. Among them, A and B can be singular or plural. The character " / " generally represents an "or" relationship between the associated objects before and after. "At least one (item) of the following" or its similar expression refers to any combination of these items, including any combination of single item (item) or plural items (items). For example, at least one (item) of a, b, or c can mean: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, c can be single or multiple.

[0169] In several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only illustrative. For example, the above division of units is only a logical function division. In actual implementation, there may be other division methods. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. The displayed or discussed coupling or direct coupling or communication connection to each other can be through some interfaces. The indirect coupling or communication connection of devices or units can be in electrical, mechanical or other forms.

[0170] The units described above as separate components may or may not be physically separated. The components displayed as units may or may not be physical units, that is, they may be located in one place, or may be distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0171] In addition, in each embodiment of this application, each functional unit can be integrated in a processing unit, or each unit can exist physically alone, or two or more units can be integrated in one unit. The above integrated unit can be implemented in the form of hardware or in the form of a software functional unit.

[0172] When the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes: various media that can store programs, such as USB flash drives, mobile hard disks, read-only memories (ROM), random access memories (RAM), magnetic disks, or optical discs.

[0173] The preferred embodiments of the embodiments of this application have been described above with reference to the accompanying drawings, which does not limit the scope of the rights of the embodiments of this application. Any modifications, equivalent replacements, and improvements made by those skilled in the art without departing from the scope and essence of the embodiments of this application shall be within the scope of the rights of the embodiments of this application.

Claims

1. A channel power spectrum estimation method, characterized in that: The method is performed by a scene-based statistical channel model, wherein the scene-based statistical channel model at least includes: a cell feature encoder, a grid feature encoder and an angle power spectrum estimation network. The method includes: Obtaining a grid position of a grid to be estimated in a target area, and obtaining a cell position and cell antenna parameters of at least one communication cell corresponding to the grid position, wherein the realistic statistical channel model is pre-trained by reference signal received powers between a plurality of grids in the target area and the communication cell; Inputting the cell location and the cell antenna parameters into the cell feature encoder for feature encoding processing to obtain a cell feature code; Inputting the grid position and the cell position into the grid feature encoder for feature encoding processing to obtain a grid feature code; The cell feature code and the grid feature code are input into the angle power spectrum estimation network to perform power spectrum estimation, so as to obtain an estimated channel power spectrum between the grid to be estimated and the communication cell.

2. The channel power spectrum estimation method according to claim 1, characterized in that: The scene-based statistical channel model further includes a converter module, which inputs the cell feature code and the grid feature code into the angle power spectrum estimation network to perform power spectrum estimation, and obtains an estimated channel power spectrum between the grid to be estimated and the communication cell, including: After performing XOR processing on the cell feature code and the grid feature code, the position embedding of the grid position is superimposed to obtain a feature processing unit; Inputting the feature processing unit into the transformer module for attention processing to obtain estimated processing data; The estimated processed data is input into the angle power spectrum estimation network to perform power spectrum estimation to obtain the estimated channel power spectrum.

3. The channel power spectrum estimation method according to claim 2, characterized in that: The converter module includes a multi-head attention module, and the feature processing unit is input into the converter module for attention processing to obtain estimated processing data, including: Acquire other feature processing units between other grids in the target area and the communication cell, and combine the other feature processing units with the feature processing unit to obtain a feature unit matrix; The feature unit matrix is ​​input into the multi-head attention module, and multi-head attention processing is performed to obtain the estimated processing data.

4. The channel power spectrum estimation method according to claim 3, characterized in that: The multi-head attention module includes a plurality of self-attention processing modules, and the feature unit matrix is ​​input into the multi-head attention module to perform multi-head attention processing to obtain the estimated processing data, including: In the multi-head attention module, a division dimension value of each attention head is obtained based on a ratio of the data dimension of the feature unit matrix to the number of attention heads; Generate a projection parameter matrix for each attention head based on the partition dimension value; In each of the self-attention processing modules, a matrix query vector, a matrix key vector and a matrix value vector are generated based on the feature unit matrix, and attention processing is performed on the matrix query vector, the matrix key vector and the matrix value vector based on the corresponding projection parameter matrix and the linear transformation matrix to obtain sub-estimation processing data; Multi-head attention processing is performed on a plurality of the sub-estimation processing data to obtain the estimation processing data.

5. The channel power spectrum estimation method according to claim 2, characterized in that: The step of inputting the estimated processed data into the angle power spectrum estimation network to perform power spectrum estimation to obtain the estimated channel power spectrum comprises: In the angle power spectrum estimation network, the estimated processed data is sequentially subjected to linear processing, activation function processing, layer normalization processing and regularization processing to obtain preliminary processed data; The preliminary processing results are sequentially subjected to residual connection enhancement processing and feature transfer processing to obtain enhanced processing data; Performing non-negative activation processing on the enhanced processed data to obtain activated processed data; Estimation processing is performed based on the activation processing data to obtain the estimated channel power spectrum.

6. The channel power spectrum estimation method according to claim 1, characterized in that: The training steps of the realistic statistical channel model include: Obtaining a channel impulse response function and a precoding matrix between each antenna in the communication cell and each of the grids; Accumulating the products of all the channel impulse response functions and the precoding matrix, squaring them, and then multiplying them by the transmit power of the communication cell to obtain a reference signal receiving power function of each grid; Performing time expectation processing on the reference signal received power function to obtain an expected reference signal received power function, wherein the expected reference signal received power function includes a coefficient matrix between the grid and the communication cell; Generate a channel power spectrum estimation model based on the coefficient matrix, the reference signal received power function, and the channel power spectrum estimation function between each of the grids and the communication cell; The realistic statistical channel model is iteratively updated multiple times based on the channel power spectrum estimation model and the multiple reference signal received powers.

7. The channel power spectrum estimation method according to claim 6, characterized in that: The iterative updating of the scene-based statistical channel model for multiple times based on the channel power spectrum estimation model includes: Obtaining a sparse penalty weight, and obtaining a penalty channel power spectrum function based on the product of the sparse penalty weight and the channel power spectrum estimation function; Generate a loss function based on the channel power spectrum estimation model and the accumulated value of the penalty channel power spectrum function; The realistic statistical channel model is iteratively updated multiple times based on the loss function and the multiple reference signal received powers.

8. A channel power spectrum estimation device, characterized in that: The method is performed by a scene-based statistical channel model, wherein the scene-based statistical channel model at least includes: a cell feature encoder, a grid feature encoder and an angle power spectrum estimation network, and the device includes: A data acquisition module, used to acquire a grid position of a grid to be estimated in a target area, and acquire a cell position and a cell antenna parameter of at least one communication cell corresponding to the grid position, wherein the realistic statistical channel model is pre-trained by a reference signal received power between a plurality of grids in the target area and the communication cell; A cell feature processing module, used for inputting the cell position and the cell antenna parameters into the cell feature encoder for feature coding processing to obtain a cell feature code; A grid feature processing module, used for inputting the grid position and the cell position into the grid feature encoder for feature encoding processing to obtain a grid feature code; The channel power spectrum estimation module is used to input the cell feature code and the grid feature code into the angle power spectrum estimation network to perform power spectrum estimation, so as to obtain the estimated channel power spectrum between the grid to be estimated and the communication cell.

9. An electronic device, characterized in that: It comprises a memory and a processor, the memory stores a computer program, and is characterized in that when the processor executes the computer program, it implements the channel power spectrum estimation method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the channel power spectrum estimation method according to any one of claims 1 to 7 is implemented.

Images