A hybrid model-driven and data-driven downlink channel estimation method and system
Patent Information
- Application Number
- CN202311706407.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-12-13
- Publication Date
- 2026-09-29
- Estimated Expiration
- 2043-12-13
AI Technical Summary
然而，由于导频的有限性，单纯模型驱动的方法或数据驱动的方法难以获得较好的下行信道估计性能，目前仍然缺乏一种有限导频开销下高性能的下行信道估计方法
[0040]有益效果：本发明提供的一种混合模型驱动与数据驱动的下行信道估计方法，通过基于模型驱动的算法初步估计角度时延域信道，利用信道维度特征构建基于多重注意力机制的角度时延域信道精细化网络来进一步细化初步信道估计，研究性能显著优于经典CS方法与已有基于神经网络的信道估计方法，通过联合利用模型驱动与数据驱动的优势来提供一种高性能的下行信道估计方法，进而提升了下行CSI获取的准确性。
Smart Images
Figure CN117834349B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of communications, specifically relating to a hybrid model-driven and data-driven downlink channel estimation method and system. Background Technology
[0002] Massive MIMO systems, as one of the key technologies for 5G and post-5G mobile communications, have been extensively researched, explored and applied in recent years. They are generally defined as systems with a large number (100 or more) of individually controllable antenna elements on the base station side of the wireless communication link, which can significantly improve system capacity, coverage, spectral efficiency and power efficiency.
[0003] To fully utilize the gains of massive MIMO systems, accurate Channel State Information (CSI) is required through channel estimation algorithms. In frequency division duplex systems, downlink CSI needs to be estimated at the user side based on the Channel State Information-Reference Signal (CSI-RS) and fed back to the base station to assist downlink transmission, such as in downlink precoding weight design. Since base stations often deploy a large number of antennas, the pilot overhead (such as CSI-RS) used for channel estimation in wireless transmission is very limited, thus posing a significant challenge to downlink channel estimation.
[0004] Since wireless channels often contain only a finite number of scatterers, the channel estimation problem in large-scale MIMO can be modeled as a compressed sensing (CS) problem, which can be solved using various classic CS-type algorithms, such as the hybrid message passing algorithm. However, CS-type algorithms have high requirements for the accuracy of model modeling and the sparsity of the channel. With the continuous development of machine learning technology, machine learning-assisted channel estimation in the field of communications has gradually become a research hotspot. However, purely data-driven methods often neglect the utilization of the model and lack certain theoretical and performance guarantees. Furthermore, due to the finite number of pilots, purely model-driven or data-driven methods are difficult to achieve good downlink channel estimation performance. Currently, there is still a lack of a high-performance downlink channel estimation method with finite pilot overhead. Summary of the Invention
[0005] Purpose of the invention: The purpose of this invention is to provide a hybrid model-driven and data-driven downlink channel estimation method and system, which improves the performance of downlink channel estimation by jointly utilizing the advantages of model-driven and data-driven approaches, and achieves accurate CSI acquisition with limited pilot overhead.
[0006] Technical Solution: To achieve the above-mentioned objectives, the present invention provides a hybrid model-driven and data-driven downlink channel estimation method, comprising the following steps:
[0007] Step 1: Transform the downlink spatial frequency domain signal reception model into the angle time delay domain signal reception model, thereby transforming the downlink spatial frequency domain channel estimation problem into the angle time delay domain sparse signal recovery problem.
[0008] Step 2: Use a model-driven compressed sensing algorithm to solve the sparse signal recovery problem and obtain a preliminary estimate of the angular delay domain channel.
[0009] Step 3: Obtain the dataset by comparing the initially estimated angle delay domain channel with the actual space-frequency domain channel;
[0010] Step 4: Based on the dimensional characteristics of the oversampled angle delay domain, construct an angle delay domain channel refinement network based on a multi-attention mechanism. The angle delay domain channel refinement network based on a multi-attention mechanism includes: multiple parallel angle attention modules constructed based on a spatial attention mechanism; and a delay attention module constructed based on a multi-head self-attention mechanism. Each angle attention module is used to process a portion of the channel's angle dimension, and the delay attention module is used to process the channel's delay dimension.
[0011] Step 5: Use the mean square error of the channel in the spatial frequency domain as the loss function to train the constructed channel refinement network in the angle delay domain.
[0012] Step 6: Use the trained angle delay domain channel refinement network to further refine the preliminary angle delay domain channel estimate output by the model-driven compressed sensing algorithm, and obtain the refined angle delay domain channel.
[0013] Step 7: Based on the refined angle delay domain channel, use two-dimensional Fourier transform to obtain the downlink spatial frequency domain channel estimation result.
[0014] In step 1, the downlink spatial frequency domain signal reception model is as follows:
[0015]
[0016] in For downlink received signals, X e For pilot matrix, For the empty frequency domain channel, This is the noise matrix.
[0017] In step 1, the angle delay domain signal reception model is as follows:
[0018]
[0019] in For the equivalent pilot matrix, From N s Point DFT matrix The middle N c Rows and the first N d The matrix extracted from the columns, N s N represents the total number of subcarriers. c N represents the number of subcarriers used for data transmission. d Indicates the number of delay taps. This represents the Kroneko product operation. and It is a phase-shifted oversampled DFT matrix. N v N h These represent the number of vertical antennas and the number of horizontal antennas, respectively. v and F h For oversampling factor, It is a sparse angular delay domain channel.
[0020] The processing steps of the model-driven compressed sensing algorithm in step 2 include: inputting the downlink received signal, the equivalent pilot matrix, and specific algorithm parameters; using the compressed sensing algorithm to perform calculations by taking advantage of the sparsity of the signal; and outputting the preliminary channel estimation results in the angle delay domain.
[0021] Step 4, which involves constructing an angle attention module based on a spatial attention mechanism, includes: constructing an angle attention layer based on a spatial attention mechanism; cascading the angle attention layer, a layer normalization layer, and a feedforward layer and adding residual connections to obtain basic angle attention blocks; and cascading an up-dimensional fully connected layer, multiple basic angle attention blocks, and a down-dimensional fully connected layer to obtain the angle attention module. The angle attention layer can be mathematically represented as follows:
[0022]
[0023]
[0024] in The angular feature map is represented by σ(·), Conv(·), AvgP(·), and MaxP(·), which represent the sigmoid activation function, 1×1×2 convolution operation, average pooling operation, and max pooling operation, respectively. d serves as the input and output of the angular attention layer, respectively. h N represents the dimension of the high-dimensional representation, · represents the tensor dot product, and N represents the high-dimensional representation dimension. Tx and N Rx These represent the number of antennas at the base station and the user end, respectively, and F is the total oversampling factor.
[0025] Step 4, which involves constructing a time-delay attention module based on a multi-head self-attention mechanism, includes: constructing a time-delay attention layer based on the multi-head self-attention mechanism; cascading the time-delay attention layer, a layer normalization layer, and a feedforward layer and adding residual connections to obtain a basic time-delay attention block; and cascading an up-dimensional fully connected layer, a position encoding module, multiple basic time-delay attention blocks, and a down-dimensional fully connected layer to obtain the time-delay attention module; wherein the time-delay attention layer can be mathematically represented as:
[0026]
[0027]
[0028] in softmax(·) represents the softmax activation function, and the superscript T indicates the transpose operation. Let N represent the i-th learnable query matrix, key matrix, value matrix, and i-th self-attention output, respectively. m Group, and has d k =d v =d h / N m , This represents the learnable projection matrix that integrates multi-head self-attention. These serve as the input and output of the time-delay attention layer, respectively.
[0029] In step 5, the loss function of the network is the mean square error of the spatial-frequency domain channel. That is, the output of the constructed angle-delay domain channel refinement network is compared with the mean square error of the real spatial-frequency domain channel using a two-dimensional Fourier transform, which can be expressed as:
[0030]
[0031] Where 2D-FT(·) and ADCRN(·) represent the two-dimensional Fourier transform and the constructed angle-delay domain channel refinement network, respectively, and are trained on the training set through supervised learning. Minimize the loss function L ad The constructed angle-delay domain channel refinement network is used to train the network, where D represents the number of samples in the dataset. These represent the tensor forms of the angle delay domain channel and the actual space-frequency domain channel, respectively, as initially estimated from the samples.
[0032] Based on the same inventive concept, the present invention also provides a hybrid model-driven and data-driven downlink channel estimation system, comprising:
[0033] The sparse channel modeling module is used to transform the downlink spatial frequency domain signal reception model into the angle time delay domain signal reception model, thereby transforming the downlink spatial frequency domain channel estimation problem into the angle time delay domain sparse signal recovery problem.
[0034] The model-driven preliminary estimation module is used to solve the problem of sparse signal recovery in the angle delay domain by using a model-driven compressed sensing algorithm to obtain a preliminary estimate of the angle delay domain channel.
[0035] The data acquisition module is used to collect the real spatial frequency domain channel and the preliminary estimated channel in the angle time delay domain as a dataset;
[0036] The network construction and training module is used to construct an angle delay domain channel refinement network based on a multi-attention mechanism, taking into account the dimensional characteristics of the oversampled angle delay domain. This network includes: multiple parallel angle attention modules constructed based on a spatial attention mechanism; and a delay attention module constructed based on a multi-head self-attention mechanism. Each angle attention module processes a portion of the channel's angle dimension, while the delay attention module processes the channel's delay dimension. The constructed angle delay domain channel refinement network is trained using the spatial frequency domain channel mean square error as the network's loss function.
[0037] It also includes a hybrid model-driven and data-driven deployment module, which takes the downlink received signal as input and uses the model-driven preliminary estimation module, the trained angle delay domain channel refinement network, and the two-dimensional Fourier transform to obtain a more accurate downlink spatial frequency domain channel estimate.
[0038] Based on the same inventive concept, the present invention also provides a computer system, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the computer program is loaded onto the processor, it implements the steps of the hybrid model-driven and data-driven downlink channel estimation method as described above.
[0039] Based on the same inventive concept, the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the hybrid model-driven and data-driven downlink channel estimation method as described above.
[0040] Beneficial effects: This invention provides a hybrid model-driven and data-driven downlink channel estimation method. It initially estimates the angle delay domain channel using a model-driven algorithm, and further refines the initial channel estimation by constructing an angle delay domain channel refinement network based on a multi-attention mechanism using channel dimension features. The research shows that the performance is significantly better than the classical CS method and existing neural network-based channel estimation methods. By jointly utilizing the advantages of model-driven and data-driven approaches, a high-performance downlink channel estimation method is provided, thereby improving the accuracy of downlink CSI acquisition. Attached Figure Description
[0041] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.
[0042] Figure 1 A flowchart of a hybrid model-driven and data-driven downlink channel estimation method provided by the present invention;
[0043] Figure 2 The flowchart illustrates the construction process of the angle delay domain channel refinement network based on a multiple attention mechanism provided in this invention.
[0044] Figure 3 This is a flowchart illustrating the specific process of the model-driven hybrid message passing algorithm used in this invention.
[0045] Figure 4 This is a schematic diagram of the angle attention layer and time delay attention layer constructed in this invention;
[0046] Figure 5 This is a structural diagram of the angle delay domain channel refinement network designed in this invention. Detailed Implementation
[0047] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the invention or its application or use. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without inventive effort are within the scope of protection of this invention.
[0048] This invention discloses a hybrid model-driven and data-driven downlink channel estimation method, applicable to scenarios where base stations are configured with large-scale antenna arrays and communicate with multiple antenna multi-user terminals, such as... Figure 1 As shown, the main steps include:
[0049] Step 1: Transform the downlink spatial frequency domain signal reception model into the angle time delay domain signal reception model, thereby transforming the downlink spatial frequency domain channel estimation problem into the angle time delay domain sparse signal recovery problem.
[0050] Step 2: Use a model-driven compressed sensing algorithm to solve the sparse signal recovery problem and obtain a preliminary estimate of the angular delay domain channel.
[0051] Step 3: Obtain the dataset by comparing the initially estimated angle delay domain channel with the actual space-frequency domain channel;
[0052] Step 4: Based on the dimensional characteristics of the oversampled angle delay domain, construct an angle delay domain channel refinement network based on a multi-attention mechanism;
[0053] Step 5: Use the mean square error of the channel in the spatial frequency domain as the loss function to train the constructed channel refinement network in the angle delay domain.
[0054] Step 6: Use the trained angle delay domain channel refinement network to further refine the preliminary angle delay domain channel estimate output by the model-driven compressed sensing algorithm, and obtain the refined angle delay domain channel.
[0055] Step 7: Based on the refined angle delay domain channel, use two-dimensional Fourier transform to obtain the downlink spatial frequency domain channel estimation result.
[0056] The following describes the detailed process of an embodiment of the present invention using a frequency division multiplexing massive MIMO-OFDM system. Its base station deployment N Tx =N po N v N h A uniform planar array of antennas to serve N Rx Antenna multi-user, N po N v N h These represent the polarization order, the number of vertical antennas, and the number of horizontal antennas, respectively. The product of the number of vertical antennas and the number of horizontal antennas is defined as N. T =N v N h The total number of subcarriers in the system is N. s , where N c Subcarriers are used for data transmission;
[0057] The downlink spatial frequency domain signal reception model in step 1 is as follows:
[0058]
[0059] in For downlink received signals, N p Indicates the number of pilot resource blocks. This is the pilot matrix, and its values are determined by the configuration of the channel state information reference signal in the 5G NR standard. For the empty frequency domain channel, Let be a noise matrix, where each element is independent and follows a zero-mean complex Gaussian distribution with variance .
[0060] Using antenna correlation and frequency correlation, the conversion relationship between the spatial frequency domain channel and the angle time delay domain channel is obtained as follows:
[0061]
[0062] in From N s Point DFT matrix The middle N c Rows and the first N d The matrix extracted from the columns, N d Indicates the number of delay taps. This represents the Kroneko product operation. and It is a phase-shifted oversampled DFT matrix. F v and F h The oversampling factor is F = Foversampling. v F h , For a sparse angle-delay domain channel, the angle-delay domain signal reception model is:
[0063]
[0064] in This is the equivalent pilot matrix.
[0065] In step 2, the model-driven compressed sensing algorithm, taking the hybrid message passing algorithm as an example, includes the following iterative processing steps: inputting the downlink received signal. Equivalent pilot matrix S e Noise variance Prior distribution of channel in angle delay domain Given the iteration number T and damping factor κ, perform T iterations and output the preliminary channel estimation result in the angle delay domain. Where η represents the prior distribution hyperparameter, E[·|·] represents the conditional expectation, and the channel elements in the angle delay domain are assumed to follow an independent and nonidentically distributed (ind) distribution. nm express The Bernoulli Gaussian distribution of the (n,m)th element can be expressed as:
[0066] p(h nm η nm ,α n )=α n p h (h nm η nm )+(1-α n )δ(hnm );
[0067] Where α n p represents the sparsity ratio. h (·) represents the Gaussian distribution during active state, η nm Let represent the corresponding Gaussian distribution parameters, then the joint prior distribution is:
[0068]
[0069] Where N = N a N d M = N po N Rx ,η={η 11 ,...,η NM The specific process of the hybrid message passing algorithm is as follows: Figure 3 As shown; it should be noted that the prior model can be assumed according to any actual scenario, including but not limited to the ind Bernoulli Gaussian prior distribution selected in this embodiment, such as the mixture Gaussian distribution, Laplace distribution, etc. The model driving part can adopt any algorithm that can provide preliminary estimation of the channel in the angle delay domain, including but not limited to the hybrid message passing algorithm selected in this embodiment, such as the orthogonal matching pursuit algorithm, sparse Bayesian learning, etc.
[0070] The tensor form of the hybrid message passing algorithm is as follows:
[0071]
[0072] Where HMP(·) represents the hybrid message passing algorithm, and These represent the downlink received signals. Preliminary channel estimation in angle delay domain The tensor form of the data is used to obtain the dataset by comparing the initially estimated angle delay domain channel with the actual space-frequency domain channel. in Represents the real spatial frequency domain channel The tensor form of the dataset, where D represents the number of samples in the dataset.
[0073] In step 4, a channel refinement network based on a multi-attention mechanism in the angle-delay domain is constructed, such as... Figure 2 As shown, it specifically includes:
[0074] Constructing an angle attention module based on spatial attention mechanism;
[0075] A time-delay attention module is constructed based on a multi-head self-attention mechanism;
[0076] The angle attention module processes the angle dimension of the channel, while the delay attention module processes the delay dimension of the channel.
[0077] An angle-delay domain channel refinement network is constructed by multiple parallel angle attention modules and one time-delay attention module.
[0078] The angle attention module constructed based on the spatial attention mechanism includes: constructing an angle attention layer based on the spatial attention mechanism; cascading the angle attention layer, layer normalization layer, and feedforward layer and adding residual connections to obtain the basic angle attention block; cascading an up-dimensional fully connected layer, N... AA (The appropriate value can be chosen based on network performance and complexity; in this example, N is used.) AA =8) basic attention blocks and a dimension-reduced fully connected layer to obtain an angle attention module; where the angle attention layer can be mathematically represented as:
[0079]
[0080]
[0081] in The angular feature map is represented by σ(·), Conv(·), AvgP(·), and MaxP(·), which represent the sigmoid activation function, 1×1×2 convolution operation, average pooling operation, and max pooling operation, respectively. d serves as the input and output of the angular attention layer, respectively. h This represents the dimension of the high-dimensional representation; in this example, d... h =512, · represents the tensor dot product.
[0082] The time-delay attention module constructed based on the multi-head self-attention mechanism includes: a time-delay attention layer constructed based on the multi-head self-attention mechanism; a cascaded time-delay attention layer, a layer normalization layer, and a feedforward layer, with residual connections to obtain a basic time-delay attention block; a cascaded up-dimensional fully connected layer; a position encoding module; and N... DA (N in this example) DA =8) delay-attention basic blocks and a dimension-reduced fully connected layer to obtain a delay-attention module; wherein the delay-attention layer can be mathematically represented as:
[0083]
[0084]
[0085] in softmax(·) represents the softmax activation function, and the superscript T indicates the transpose operation. Let N represent the i-th learnable query matrix, key matrix, value matrix, and i-th self-attention output, respectively. mGroup, and has d k =d v =d h / N m , This represents the learnable projection matrix that integrates multi-head self-attention. These serve as the input and output of the time-delay attention layer, respectively.
[0086] The structures of the angle attention layer based on the spatial attention mechanism and the time delay attention layer based on the multi-head self-attention mechanism are as follows: Figure 4 As shown.
[0087] The angle delay domain channel refinement network consists of F angle attention modules that process the channel angle dimension and one delay attention module that processes the channel delay dimension, performing preliminary estimation of the angle delay domain channel. The virtual and real parts are extracted and input into the network in two dimensions. The network outputs a refined angle delay domain channel. Similarly, its structure is as follows Figure 5 As shown, R represents tensor shape transformation operation and C represents tensor stacking operation.
[0088] In step 5, the loss function of the network is the mean square error of the spatial-frequency domain channel. That is, the output of the constructed angle-delay domain channel refinement network is compared with the mean square error of the real spatial-frequency domain channel using a two-dimensional Fourier transform, which can be expressed as:
[0089]
[0090] Where 2D-FT(·) and ADCRN(·) represent the two-dimensional Fourier transform and the constructed angle-delay domain channel refinement network, respectively, and are trained on the training set through supervised learning. Minimize the loss function L ad To train the constructed angle delay domain channel refinement network.
[0091] The deployment of the downlink channel estimation method includes: obtaining a preliminary angle delay domain channel estimate using a model-driven hybrid message passing algorithm; further refining the preliminary angle delay domain channel estimate output by the hybrid message passing algorithm using a data-driven trained angle delay domain channel refinement network to obtain a refined angle delay domain channel; and finally, obtaining the downlink full-band channel estimation result using a two-dimensional Fourier transform based on the refined angle delay domain channel. The entire process can be mathematically represented as follows:
[0092]
[0093]
[0094]
[0095] in and These represent the refined angle delay domain channel and the obtained downlink full-band channel estimation results, respectively.
[0096] Based on the same inventive concept, this embodiment of the invention also provides a hybrid model-driven and data-driven downlink channel estimation system, including: a sparse channel modeling module, used to transform the downlink spatial frequency domain signal reception model into an angle time delay domain signal reception model, thereby transforming the downlink spatial frequency domain channel estimation problem into an angle time delay domain sparse signal recovery problem.
[0097] The model-driven preliminary estimation module is used to solve the problem of sparse signal recovery in the angle delay domain by using a model-driven compressed sensing algorithm to obtain a preliminary estimate of the angle delay domain channel.
[0098] The data acquisition module is used to collect the real spatial frequency domain channel and the preliminary estimated channel in the angle time delay domain as a dataset;
[0099] The network construction and training module is used to construct an angle delay domain channel refinement network based on a multi-attention mechanism, taking into account the dimensional characteristics of the oversampled angle delay domain. This network includes: multiple parallel angle attention modules constructed based on a spatial attention mechanism; and a delay attention module constructed based on a multi-head self-attention mechanism. Each angle attention module processes a portion of the channel's angle dimension, while the delay attention module processes the channel's delay dimension. The constructed angle delay domain channel refinement network is trained using the spatial frequency domain channel mean square error as the network's loss function.
[0100] The system also includes a hybrid model-driven and data-driven deployment module, which uses the downlink received signal as input and employs a model-driven preliminary estimation module, a trained angle delay domain channel refinement network, and a two-dimensional Fourier transform to obtain a more accurate downlink spatial-frequency domain channel estimate. Specific implementation details for each module are provided in the above method embodiments and will not be repeated here.
[0101] On the other hand, embodiments of the present invention also provide a computer system, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the computer program is loaded onto the processor, it implements the steps of hybrid model-driven and data-driven downlink channel estimation as described in any of the above embodiments.
[0102] On the other hand, embodiments of the present invention also provide a computer-readable storage medium storing a computer program that, when loaded onto a processor, implements the steps of hybrid model-driven and data-driven downlink channel estimation as described above.
[0103] In this embodiment of the invention, a model-driven hybrid message passing algorithm is employed, which utilizes the sparsity of the channel and prior knowledge of the probability distribution. A data-driven angle delay domain channel refinement network is used to extract potential features of the angle delay domain channel through big data. Downlink channel estimation is completed in a joint model-driven and data-driven manner. High-performance downlink channel estimation is achieved in a frequency division duplex large-scale MIMO-OFDM system with limited pilot overhead, thereby improving the accuracy of downlink CSI acquisition.
[0104] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
Claims
1. A hybrid model-driven and data-driven downlink channel estimation method, characterized in that, Includes the following steps: Step 1: Transform the downlink spatial frequency domain signal reception model into the angle time delay domain signal reception model, thereby transforming the downlink spatial frequency domain channel estimation problem into the angle time delay domain sparse signal recovery problem. Step 2: Use a model-driven compressed sensing algorithm to solve the sparse signal recovery problem and obtain a preliminary estimate of the angular delay domain channel. Step 3: Obtain the dataset by comparing the initially estimated angle delay domain channel with the actual space-frequency domain channel; Step 4: Based on the dimensional characteristics of the oversampled angle delay domain, construct an angle delay domain channel refinement network based on a multi-attention mechanism. The angle delay domain channel refinement network based on a multi-attention mechanism includes: multiple parallel angle attention modules constructed based on a spatial attention mechanism; and a delay attention module constructed based on a multi-head self-attention mechanism. Each angle attention module is used to process a portion of the channel's angle dimension, and the delay attention module is used to process the channel's delay dimension. Step 5: Use the mean square error of the channel in the spatial frequency domain as the loss function to train the constructed channel refinement network in the angle delay domain. Step 6: Use the trained angle delay domain channel refinement network to further refine the preliminary angle delay domain channel estimate output by the model-driven compressed sensing algorithm, and obtain the refined angle delay domain channel. Step 7: Based on the refined angle delay domain channel, use two-dimensional Fourier transform to obtain the downlink spatial frequency domain channel estimation result.
2. The downlink channel estimation method based on a hybrid model-driven and data-driven approach according to claim 1, characterized in that, In step 1, the downlink spatial frequency domain signal reception model is as follows: in For downlink received signals, X e For pilot matrix, For the empty frequency domain channel, This is the noise matrix.
3. The downlink channel estimation method based on a hybrid model-driven and data-driven approach according to claim 1, characterized in that, In step 1, the angle delay domain signal reception model is as follows: in For downlink received signals, For the equivalent pilot matrix, X e For pilot matrix, From N s Point DFT matrix The middle N c Rows and the first N d The matrix extracted from the columns, N s N represents the total number of subcarriers. c N represents the number of subcarriers used for data transmission. d Indicates the number of delay taps. This represents the Kroneko product operation. and It is a phase-shifted oversampled DFT matrix. N v N h These represent the number of vertical antennas and the number of horizontal antennas, respectively. v and F h For oversampling factor, For sparse angular delay domain channels, This is the noise matrix.
4. The downlink channel estimation method based on a hybrid model-driven and data-driven approach according to claim 1, characterized in that, The processing steps of the model-driven compressed sensing algorithm in step 2 include: inputting the downlink received signal, the equivalent pilot matrix, and specific algorithm parameters; using the compressed sensing algorithm to perform calculations by taking advantage of the sparsity of the signal; and outputting the preliminary channel estimation results in the angle delay domain.
5. The downlink channel estimation method based on a hybrid model-driven and data-driven approach according to claim 1, characterized in that, Step 4, which involves constructing an angle attention module based on a spatial attention mechanism, includes: constructing an angle attention layer based on a spatial attention mechanism; cascading the angle attention layer, a layer normalization layer, and a feedforward layer and adding residual connections to obtain basic angle attention blocks; and cascading an up-dimensional fully connected layer, multiple basic angle attention blocks, and a down-dimensional fully connected layer to obtain the angle attention module; wherein the angle attention layer is represented as follows: in The angular feature map is represented by σ(·), Conv(·), AvgP(·), and MaxP(·), which represent the sigmoid activation function, 1×1×2 convolution operation, average pooling operation, and max pooling operation, respectively. d serves as the input and output of the angular attention layer, respectively. h N represents the dimension of the high-dimensional representation, · represents the tensor dot product, and N represents the high-dimensional representation dimension. Tx and N Rx These represent the number of antennas at the base station and the user end, respectively, and F is the total oversampling factor.
6. The downlink channel estimation method based on a hybrid model-driven and data-driven approach according to claim 1, characterized in that, Step 4, which involves constructing a time-delay attention module based on a multi-head self-attention mechanism, includes: constructing a time-delay attention layer based on the multi-head self-attention mechanism; cascading the time-delay attention layer, a layer normalization layer, and a feedforward layer and adding residual connections to obtain a basic time-delay attention block; and cascading an up-dimensional fully connected layer, a position encoding module, multiple basic time-delay attention blocks, and a down-dimensional fully connected layer to obtain the time-delay attention module; wherein the time-delay attention layer is represented as follows: in softmax(·) represents the softmax activation function, and the superscript T indicates the transpose operation. Let N represent the i-th learnable query matrix, key matrix, value matrix, and i-th self-attention output, respectively. m Group, and has d k =d v =d h / N m , This represents the learnable projection matrix that integrates multi-head self-attention. d serves as the input and output of the time-delay attention layer, respectively. h N represents the dimension of the high-dimensional representation. d Indicates the number of delay taps.
7. The downlink channel estimation method based on a hybrid model-driven and data-driven approach according to claim 1, characterized in that, In step 5, the loss function of the network is the mean square error of the spatial-frequency domain channel. That is, the output of the constructed angle-delay domain channel refinement network is used to calculate the mean square error with the real spatial-frequency domain channel through a two-dimensional Fourier transform, expressed as: Where 2D-FT(·) and ADCRN(·) represent the two-dimensional Fourier transform and the constructed angle-delay domain channel refinement network, respectively, and are trained on the training set through supervised learning. Minimize the loss function L ad The constructed angle-delay domain channel refinement network is used to train the network, where D represents the number of samples in the dataset. These represent the tensor forms of the angle delay domain channel and the actual space-frequency domain channel, respectively, as initially estimated from the samples.
8. A hybrid model-driven and data-driven downlink channel estimation system, characterized in that, include: The sparse channel modeling module is used to transform the downlink spatial frequency domain signal reception model into the angle time delay domain signal reception model, thereby transforming the downlink spatial frequency domain channel estimation problem into the angle time delay domain sparse signal recovery problem. The model-driven preliminary estimation module is used to solve the problem of sparse signal recovery in the angle delay domain by using a model-driven compressed sensing algorithm to obtain a preliminary estimate of the angle delay domain channel. The data acquisition module is used to collect the real spatial frequency domain channel and the preliminary estimated channel in the angle time delay domain as a dataset; The network construction and training module is used to construct an angle delay domain channel refinement network based on a multi-attention mechanism, taking into account the dimensional characteristics of the oversampled angle delay domain. This network includes: multiple parallel angle attention modules constructed based on a spatial attention mechanism; and a delay attention module constructed based on a multi-head self-attention mechanism. Each angle attention module processes a portion of the channel's angle dimension, while the delay attention module processes the channel's delay dimension. The constructed angle delay domain channel refinement network is trained using the spatial frequency domain channel mean square error as the network's loss function. It also includes a hybrid model-driven and data-driven deployment module, which takes the downlink received signal as input and uses the model-driven preliminary estimation module, the trained angle delay domain channel refinement network, and the two-dimensional Fourier transform to obtain a more accurate downlink spatial frequency domain channel estimate.
9. A computer system comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the computer program is loaded into the processor, it implements the steps of the hybrid model-driven and data-driven downlink channel estimation method according to any one of claims 1-7.
10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the hybrid model-driven and data-driven downlink channel estimation method according to any one of claims 1-7.