Channel estimation method, system, equipment, medium and terminal
Through the combination of deep learning compression perception technology, the problems of high computational complexity, insufficient accuracy and insufficient robustness of channel estimation in wireless communications are solved, and robust channel estimation in large dynamic signal-to-noise ratio range are realized , suitable for a variety of channel environments.
Patent Information
- Application Number
- CN202411125746.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-16
- Publication Date
- 2024-11-15
AI Technical Summary
The prior art channel estimation method in wireless communications has problems such as high computational complexity, insufficient estimation accuracy under low signal-to-noise ratio, and lack of robustness and adaptability, especially in large-scale antenna array systems and user mobility scenarios. High performance channel estimation.
Using deep learning-based compression perception technology, compressed sampling, linear preliminary reconstruction and nonlinear deep reconstruction of channel images are realized through the combination of sampling sub-network, preliminary reconstruction sub-network and deep reconstruction sub-network, reducing the computational complexity and improving channel estimation. Accuracy and robustness.
Achieve robust channel estimation performance in a large dynamic signal-to-noise ratio range, reduces the computational complexity of channel estimation, improves channel estimation accuracy and robustness in low signal-to-noise ratio environments, and is suitable for time frequency, antenna and beam Channel estimation in the domain has broad applicability and flexibility.
Smart Images

Figure CN118972012A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to but is not limited to the field of wireless communication technology, and in particular relates to a channel estimation method, system, device, medium and terminal. Background Art
[0002] From the 4G LTE standard (R8) released in 2009 to the 5G NR standard (R15) released in 2017, the 3GPP wireless communication standardization organization has been continuously improving and enhancing the ability to obtain channel information. The quality of channel estimation and channel information directly determines the overall performance of the wireless communication system and is the key to the success or failure of the wireless communication system. With the increasing application of large-scale antenna arrays, higher frequencies, wider bandwidths, and beam-based communications, the effective operation of the next-generation 6G wireless communication system is increasingly dependent on the high-precision acquisition of channel information.
[0003] Traditional channel estimation:
[0004] Taking the traditional single-input single-output OFDM system as an example, the received signal is
[0005] y=XFg+n (1)
[0006] Where X is the pilot signal, which is an N×N diagonal matrix due to mutual orthogonality, F is an N×N DFT matrix, n is an N×1 noise vector, g is an N×1 channel time domain impulse response vector, and its element g k for
[0007]
[0008] where k∈[0,N-1], τ l is the multipath delay, when When it is not an integer, spectrum leakage will occur. Define h = Fg as the frequency domain impulse response of the channel, and the least square channel estimation LS (Least Square) can be expressed as
[0009]
[0010] LS channel estimation does not require channel statistics information and is simple and convenient to implement, so it has been widely used in the industry. The disadvantage of LS channel estimation is that it amplifies the noise energy, especially when the channel is deeply faded, which will cause serious performance degradation. Therefore, the minimum mean square error coefficient MMSE (Minimum Mean Square Error) is introduced to improve the performance of LS channel estimation. The linear minimum mean square error channel estimation LMMSE can be expressed as
[0011]
[0012] where R hh = E{hh H} is the autocorrelation matrix of the channel, and β is a constant determined by the constellation points (for 16-QAM, β = 17 / 9). LMMSE can greatly improve the channel estimation performance, but its computational and implementation complexity also increases significantly. To reduce the computational complexity of LMMSE channel estimation, researchers use the time-domain sparsity of wireless signals to approximately implement LMMSE. For example, in reference [1], only the first L (1 < L < N) elements of g are considered. In its cross-correlation matrix R gg , only the first L elements are retained, and the other elements are set to zero, thereby reducing the computational complexity. The time-domain sparsity of the signal is also reflected in the eigenvalue space. Performing SVD singular value decomposition on R hh gives R hh = UΛU H . Only the eigenvalue vectors of rank p are retained, and thus the low-rank channel estimate is
[0013]
[0014] where Δ p is a diagonal matrix, and the first p diagonal elements δ k = λ k / (λ k + β / SNR) (λ k are the eigenvalues of R hh ), and the elements in other positions are zero. The low-rank channel estimate can approximate the theoretical optimal performance, but it still requires a singular value decomposition operation with high computational complexity. On the other hand, both LMMSE and low-rank channel estimation are based on the known channel statistical information R hh . In practice, it is very difficult and inefficient to obtain the channel statistical information R hh . It is necessary to rely on long-term channel statistical information to obtain a sufficiently accurate approximation of R hh , making the implementation complexity and cost of algorithms of LMMSE and its low-rank channel estimation types still very high.
[0015] With the help of further exploration of signal sparsity, orthogonal matching pursuit OMP (Orthogonal Matching Pursuit) and basis pursuit BP (Basis Pursuit) get rid of the dependence on channel statistical information. As a result, the traditional channel estimation problem is transformed into a sparse recovery and reconstruction problem. Traditional time-domain channel estimation only collects the energy delay characteristics PDP (Power Delay Profile) of the channel, and estimates the delay corresponding to the multipath peak point from the obtained PDP, and then reconstructs the channel according to the position corresponding to the main energy delay. With the increase in the number of base station antennas, array signal processing comes into play. Taking a linear antenna array ULA (Uniform Linear Array) equipped with M antenna units as an example, the channel model between it and a single-antenna user can be expressed as
[0016]
[0017] Where a0u(θ0) is the line-of-sight (LOS) component, a0 is the path gain, the component corresponding to 1≤p≤P is the non-line-of-sight (NLOS) path component, P is the total number of multipaths, and θ p is the angle of arrival AoA (Angle of Arrive) of path p, u(θ p ) is usually called the steering vector of the antenna array, and its elements can be expressed as m∈[q-(M-1) / 2,q=0,1,…,M-1],λ is the wavelength, d is the antenna spacing, and d=λ / 2 is usually satisfied. In a typical 3GPP environment, the angular spread is relatively limited, generally between 2°-5°. Similar to the sparsity in the time domain, the wireless signal also exhibits spatial sparsity. Therefore, the channel estimation problem is transformed into the spatial orientation estimation of the main path. The estimation of the arrival angle is a classic problem in array signal processing, such as the MUSIC and ESPRIT algorithms. These classic algorithms for array signal processing were originally designed for radar perception. They are based on the assumption of blind estimation and do not have pre-designed training sequences like commercial wireless communication systems. Moreover, these algorithms are highly complex and difficult to apply to commercial wireless communication systems. Similarly, the prior art has proposed a series of spatial low-rank channel estimation algorithms. The basic idea is to use fast Fourier transform to transform the received signal into the beam domain, find the beam direction containing the main energy, set the beams in other directions to zero, and then transform the zeroed channel back through inverse Fourier transform, thereby realizing rapid estimation and reconstruction of the wireless channel.
[0018] These beam-domain low-rank channel estimates use the spatial sparsity of wireless signals to reduce the dimensionality of the originally large channels, reducing the demand for RF resources in large-scale antenna array systems, while also effectively alleviating the data pressure of the fronthaul between the RF end and the digital end of the edge antenna. However, its defects are also obvious. The beam-domain low-rank channel estimation only considers spatial sparsity, while wireless signals do not exhibit spatial sparsity in all scenarios. Secondly, in scenarios where wireless signals exhibit spatial sparsity, their main paths are often too close to be distinguished. Furthermore, the Fourier transform commonly used in these beam-domain channel estimates does not have a good enough ability to distinguish, and the sidelobe beams can cause serious interference.
[0019] Channel Estimation Based on Deep Learning:
[0020] Image denoising is a classic and still active topic in image processing. It requires removing noise, errors or disturbances in the signal as much as possible, that is, recovering a clean image X from a noisy observed image Y=X+N, where N is some form of noise, generally additive Gaussian white noise. Image denoising based on deep learning has been successfully applied to the field of image processing in the past few years. Based on the residual network, an image denoising neural network was constructed and named the deep denoising network DnCNN, which can be applied to low-level image tasks such as image denoising and super-resolution. DnCNN requires training different networks for images with different noise levels to obtain accurate results, which lacks flexibility. Therefore, the researchers have developed a fast and flexible denoising network FFDNet. The advantage of this network is that it can input the known noise level as a feature map into the denoising network, so that a single network can handle different levels of noise. The prior art has further proposed CBDNet, designed a noise level estimation subnetwork to estimate the unknown noise intensity, and then input the noise level feature map and the noisy image into the U-Net network for denoising. The network architecture of U-Net is as follows Figure 1 As shown in the figure, each square represents a fully connected convolutional layer, and the gray dotted arrows represent skip connections. The U-Net network architecture is divided into two parts: the left half is the encoder (Encoder), and the right half is the decoder (Decoder). The encoder is used to compress noisy images, and the decoder restores the image by upsampling with padding twice. The skip connection combines the feature map obtained in the encoding process with the feature map in the decoding process for feature enhancement. Finally, a convolutional layer with a kernel size of 1×1 is used for dimensional mapping to output the denoised image. Since the entire network architecture is in the shape of a capital U, it is named U-Net.
[0021] Image denoising and channel estimation have different application scenarios, but from a mathematical point of view, channel estimation is equivalent to image denoising, both of which are to recover clean data from a contaminated observation data. As wireless communication systems transition from linear antenna arrays (ULA) to planar antenna arrays (UPA), the channel matrix expands from a one-dimensional vector to a multi-dimensional tensor. Moreover, the channel signals of large-scale antenna arrays are sparse in time-frequency-space dimensions, which gives the channel signals image-like properties. The channel signals are also a type of image, which the present invention calls channel images. The development of large-scale MIMO systems has inspired researchers to apply convolutional neural networks, which have achieved excellent performance in image processing, to channel estimation.
[0022] In recent years, channel estimation based on convolutional neural networks has achieved remarkable results in large-scale antenna array systems. According to the implementation method of network training, channel estimation based on convolutional neural networks can be roughly divided into two categories: model-driven and data-driven. Model-driven systems are based on predefined mathematical models and their optimization, while data-driven systems are based on data and their learning. Model-driven methods assume that the randomness of wireless channels follows a certain mathematical distribution, such as modeling the beam domain channel as a Gaussian mixture GM (Gaussian mixture) distribution, and deriving the approximate message passing network GMLAMP (GM learned approximate message passing) based on Gaussian mixture priors. In contrast, data-driven methods do not rely on mathematical models of wireless channel distribution and expert knowledge, but use collected channel data to train neural networks. Data-driven deep learning algorithms are gaining more and more attention in the field of channel estimation due to their powerful function fitting capabilities and generalization capabilities without the need for preset models. Data-driven channel estimation, drawing on the application of deep learning in image denoising, regards the received channel signal as an image containing noise, and uses deep learning to denoise the channel image, achieving performance that is superior to traditional LMMSE channel estimation. The target detection network YOLO (You Only Look Once) is used for channel estimation. By reconstructing the channel data into an image, the YOLO algorithm is used to detect the bright spots in the image, thereby obtaining perception information such as angle and delay, and then reconstructing the channel. It is proposed to apply the multi-layer perceptron architecture to large-scale MIMO systems, learn channel features through data-driven learning, and achieve better performance than traditional channel estimation methods.
[0023] The super-resolution problem and channel estimation in the field of image processing also have many similarities. First, both involve reconstructing missing information from incomplete or blurred observation data. Channel estimation attempts to infer the channel information at the data from the channel information at the pilot position. Super-resolution attempts to reconstruct high-resolution image details from low-resolution images. Moreover, in image super-resolution, the process of obtaining a low-resolution image from a high-resolution image is generally modeled as a spatially equally spaced downsampling process, while in commercial wireless communication systems, pilots are placed at equal intervals on space-time-frequency resources. Second, both rely on the correlation of channel or image models. In channel estimation, the channel changes according to a certain rule, and the channel matrix or tensor has a certain sparsity; in image super-resolution, the image has certain statistical features, and these features are used to infer the details of the high-resolution image. In the MIMO-OFDM system, the image super-resolution neural network SRCNN and EDSR are used for channel estimation interpolation, and the network trained with the statistical channel model is applied to the actual WiFi system, and the results are better than the traditional linear interpolation and Gaussian interpolation methods. Combining image super-resolution and image denoising techniques, it is proposed to cascade the super-resolution network SRCNN and the denoising network DnCNN to obtain a new network for channel estimation, called ChannelNet. Figure 2 ChannelNet adopts a two-stage approach during training, treating the time-frequency channel matrix of the OFDM system as a two-dimensional image, using the channel matrix at the pilot as a low-resolution input, and outputting a complete high-resolution time-frequency channel matrix.
[0024] In view of the above analysis, the technical problems that need to be urgently solved in the existing technology are: the convolutional neural network model derived from image processing is directly applied to the perception and estimation of wireless channels, which has achieved performance gains, but this advantage is not necessarily universal. For example, in the field of image processing, the image signal-to-noise ratio is usually relatively fixed and has a small dynamic range. In the field of wireless communications, due to the mobility of users, the signal-to-noise ratio of channel images usually changes dynamically over a large range. Summary of the invention
[0025] In view of the problems existing in the prior art, the present invention provides a channel estimation method, system, device, medium and terminal.
[0026] The present invention is implemented in this way: a channel estimation method comprises:
[0027] S1, the sampling subnetwork compresses and samples the input noisy channel image to obtain the compressed sensing measurement value;
[0028] S2, the two reconstruction sub-networks try to learn an end-to-end mapping from channel measurements to the true channel;
[0029] S3, the preliminary reconstruction subnetwork performs a linear preliminary reconstruction of the channel measurement values;
[0030] S4 then performs nonlinear deep reconstruction through the deep reconstruction sub-network.
[0031] Furthermore, the convolutional layer is used to simulate and replace the compressed sensing sampling, and the compressed sensing measurement value is obtained as the initial feature map of the channel;
[0032] The mapping relationship between the compression sense measurement value and the real channel is obtained through convolution layer training, and the weight coefficient of the mapping relationship is obtained;
[0033] The weight coefficient of the mapping relationship between the trained compressed sensing measurement value and the real channel is used for online channel estimation, and a preliminary reconstructed channel is obtained according to the input compressed sensing measurement value;
[0034] The weight coefficients of the mapping relationship between the preliminary reconstructed channel and the real channel are obtained through deep convolutional residual network training;
[0035] The weight coefficient of the mapping relationship between the trained preliminary reconstructed channel and the real channel is used for online channel estimation, and the deep reconstructed channel is obtained according to the input preliminary reconstructed channel;
[0036] The new compressed sensing channel estimation network can estimate the wireless channel in the following domains: time domain, frequency domain, antenna domain, and beam domain.
[0037] Furthermore, the sampling subnetwork specifically includes:
[0038] In order to reduce the complexity of sampling, the compressed sampling subnetwork adopts block-wise compressed sampling, which divides the input two-channel image into (M / B)×(K / B) non-overlapping image blocks with a block size of B×B×2. B can be adjusted and needs to be divisible by both M and K. Then, block-wise sampling is implemented through a convolutional layer, where n is used. b A B×B×2 convolution filter, the stride of the convolution kernel is set to B×B so that it acts on different blocks when sliding, and no zero padding is used. v Represents the number of compressed sensing measurements obtained for each block. When the sampling rate is set to δ, n v =δ·2B 2 In fact, this convolutional layer simulates the sampling process of traditional block compressed sensing, where each block uses the same sampling matrix:
[0039] y b =Φ b x b (9)
[0040] in is the vectorized image block, is the block sampling matrix, is the compressed sensing measurement value of the image block.
[0041] Furthermore, the preliminary reconstruction sub-network specifically includes:
[0042] The purpose of the preliminary reconstruction network is to preliminarily reconstruct each image block from the compressed sensing measurements through a convolutional layer. After passing through the sampling subnetwork, all the output measurements form a (M / B)×(K / B)×n b A tensor where the measurement value of each block is a 1×1×n b The present invention uses 2B 2 1×1×n b The convolution filter is used to reconstruct all blocks, and the convolution kernel step size is 1×1. Then, the size of the output tensor is changed from (M / B)×(K / B)×2B through the "Reshape+concatenate" function. 2 Reshape to M×K×2 to obtain the preliminary reconstructed channel image. The convolution layer of the preliminary reconstruction network simulates the following matrix operation:
[0043]
[0044] in is the initial reconstructed image block after vectorization, is the initial reconstruction matrix. Φ b and The first two sub-networks simulate linear operations, so no activation function is used.
[0045] Furthermore, the deep reconstruction sub-network specifically includes:
[0046] The purpose of the deep reconstruction subnetwork is to further refine the channel image and improve the reconstruction quality. The deep reconstruction subnetwork contains a total of 15 convolutional layers. First, the initially reconstructed M×K×2 channel image is input into the first convolutional layer for feature extraction to generate 64 feature maps, which then pass through 6 cascaded residual blocks and finally pass through two ordinary convolutional layers (the first of which uses the ReLU function) to form the final M×K×2 reconstructed channel image. The residual block contains two batch normalization (BN) layers, a ReLU function and a dilated convolution layer, as well as a short skip connection, such as Figure 3As shown in the figure. The skip connection used in the residual block promotes gradient propagation without losing local information, thus reducing the difficulty of training. The dilation rate used in the dilated convolution is 4. The traditional convolution operation only considers the adjacent input pixels at each position, while the dilated convolution introduces holes in the convolution kernel, which can expand the receptive field while keeping the number of parameters unchanged, so that the network can utilize a wider range of spatial information. In the deep reconstruction subnetwork, the spatial size of the convolution kernel is 3×3, and the number of feature channels in the middle layer is 64.
[0047] Another object of the present invention is to provide a channel estimation system for implementing the channel estimation method, comprising:
[0048] A sampling subnetwork module is used to compress and sample the input noisy channel image to obtain a compressed sensing measurement value;
[0049] A preliminary reconstruction subnetwork module is used for performing a linear preliminary reconstruction of the channel measurement values;
[0050] Deep reconstruction sub-network module, used for nonlinear deep reconstruction.
[0051] Another object of the present invention is to provide a computer device, the computer device includes a memory and a processor, the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the channel estimation method.
[0052] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, enables the processor to perform the steps of the channel estimation method.
[0053] Another object of the present invention is to provide an information data processing terminal, which includes the channel estimation system.
[0054] In combination with the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solutions to be protected by the present invention are as follows:
[0055] First, the channel estimation network CSCENet based on deep compressed sensing proposed in the present invention can achieve robust performance only in the time-frequency domain, with a large SNR dynamic range and in scenarios where users move, avoiding additional hardware (3D lens antenna array) or software (beam domain spatial transformation) costs.
[0056] The present invention establishes a channel model according to 3GPP TR 38.901 for simulation. The frequency band used is 5GHz. CSCENet performs well in a large dynamic SNR range of [-20dB 10dB]. When the SNR is -15dB, CSCENet achieves nearly 21dB and 7dB gains compared with LS and LMMSE, respectively. For high SNR scenarios, CSCENet can achieve almost the same performance as LMMSE (SNR from 5 to 10dB).
[0057] In addition to the high online computational complexity, LMMSE also requires prior statistical information of the channel, which often requires considerable resources to obtain in practice. Therefore, LMMSE is usually only used as a benchmark for measuring the quality of channel estimation in theory, but is not feasible in practice. The complexity and performance of CSCENet are separated. The most complex part is completed by offline training, and the weight coefficients are obtained and applied to the online reasoning stage. In this way, online reasoning is just a simple addition, subtraction, multiplication and division operation. The online calculation complexity is much smaller than LMMSE, and there is no need to obtain prior knowledge of the channel. After beam transformation, the channel will be sparser, and CSCENet will achieve better performance.
[0058] Second, as auxiliary evidence of the inventiveness of the claims of the present invention, it is also reflected in the following important aspects:
[0059] (1) The expected benefits and commercial value of the technical solution of the present invention after transformation are: enabling the actual communication system to achieve the optimal performance comparable to the traditional LMMSE channel estimation theory with extremely low implementation complexity and cost. What is even more valuable is that in low signal-to-noise ratio scenarios with poor channel conditions, the channel estimation performance that exceeds LMMSE is achieved. On the one hand, it makes the communication system more robust, and on the other hand, it enables the communication system to replace the existing low-complexity and low-performance traditional channel estimation on a large scale at low cost, while achieving high performance with the help of CSCENet, which has extremely high commercial value.
[0060] (2) The technical solution of the present invention overcomes technical prejudice: For a long time, the field of wireless communication has not been optimistic about the application of artificial intelligence in the physical layer. The reason is that artificial intelligence represented by deep learning is trained for specific scenarios and is not universal. The open environment of wireless communication systems and the mobility of users determine the diversity of their scenarios. The CECENet proposed in the present invention shows robust channel estimation performance and strong generalization ability in a large SNR dynamic range and in scenarios where users move, overcoming the long-standing technical prejudice and providing a solution for wireless physical layer applications based on artificial intelligence.
[0061] Third, the technical problems of the prior art solved by the technical solution of the present invention
[0062] 1) High computational complexity problem:
[0063] Problems with existing technologies: Traditional channel estimation methods usually need to process a large amount of sampled data when performing accurate channel estimation, which leads to high computational complexity. Especially in multi-antenna systems or frequency domain and time domain channel estimation, the computational burden is very heavy, making it difficult to achieve real-time processing in actual communication systems.
[0064] The solution of the present invention is to introduce ultra-compressed sensing technology and use a sampling sub-network to compress and sample the channel image, which significantly reduces the amount of data to be processed, thereby reducing the computational complexity of channel estimation. This technology enables channel estimation to be processed quickly with limited computing resources.
[0065] 2) Insufficient estimation accuracy caused by low signal-to-noise ratio:
[0066] Existing technical problems: In low signal-to-noise ratio environments, traditional channel estimation methods are easily affected by noise, resulting in reduced estimation accuracy. This is particularly prominent in actual wireless communication systems, directly affecting the stability and reliability of communications.
[0067] The solution of the present invention: The present invention adopts the deep convolution residual network in deep learning, and enhances the feature extraction and reconstruction capabilities of channel images through the design of multiple convolution layers and residual blocks. In particular, the receptive field is expanded by dilated convolution, so that the network can capture a wider range of channel features and improve the channel estimation accuracy in low signal-to-noise ratio environments.
[0068] 3) Robustness and adaptability of channel estimation:
[0069] Problems with existing technologies: Traditional channel estimation methods are usually effective for specific channel models and lack robustness and adaptability to different channel environments. When the channel environment changes, the performance of existing methods drops significantly.
[0070] The solution of the present invention: Through the design of two reconstruction sub-networks, the technical solution of the present invention can not only process linear channel characteristics, but also capture nonlinear channel characteristics through the deep reconstruction sub-network. This combined reconstruction method greatly improves the adaptability to a variety of channel environments, making the channel estimation more robust and able to cope with more complex wireless channel environments.
[0071] The significant technical progress obtained by the present invention:
[0072] 1) Significantly reduce computational complexity:
[0073] Through ultra-compressed sensing technology and the design of sampling subnetworks, the present invention significantly reduces the amount of sampling data required for channel estimation. This not only reduces the computational complexity, but also enables channel estimation to be processed in real time on resource-constrained devices, which is difficult to achieve with traditional methods.
[0074] ) Improve channel estimation accuracy:
[0075] The present invention greatly improves the accuracy of channel estimation by introducing deep convolutional residual network and dilated convolution technology. In particular, it can accurately reconstruct the real channel characteristics in a low signal-to-noise ratio environment, overcoming the problem that traditional methods perform poorly in low signal-to-noise ratio environments.
[0076] 3) Enhance system robustness and adaptability:
[0077] By adopting two reconstruction subnetworks to process linear and nonlinear channel characteristics respectively, the present invention can adapt to various complex wireless channel environments, significantly improves the robustness and adaptability of channel estimation, and is suitable for more practical application scenarios.
[0078] 4) Expand application areas:
[0079] The technical solution of the present invention is not only applicable to traditional time domain and frequency domain channel estimation, but can also be extended to channel estimation in antenna domain and beam domain. It has wide applicability and flexibility and can meet the development needs of future wireless communication technologies.
[0080] The present invention solves many key problems in the prior art through innovative technical solutions, and has made significant technical progress in terms of computational complexity, estimation accuracy, system robustness, etc. These advances not only improve the performance of channel estimation, but also expand its application scope, making the present invention have important application value in actual communication systems. BRIEF DESCRIPTION OF THE DRAWINGS
[0081] Figure 1 Schematic diagram of the network architecture of the U-Net model provided by an embodiment of the present invention;
[0082] Figure 2 Schematic diagram of a ChannelNet super-resolution channel estimation solution provided by an embodiment of the present invention;
[0083] Figure 3 is a flow chart of a channel estimation method provided by an embodiment of the present invention;
[0084] Figure 4 Schematic diagram of a channel estimation network structure based on deep compressed sensing provided by an embodiment of the present invention;
[0085] Figure 5This is a performance diagram when the user speed is 3 km / h provided by an embodiment of the present invention;
[0086] Figure 6 It is a performance diagram provided by an embodiment of the present invention when the user speed is 30 km / h. DETAILED DESCRIPTION
[0087] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.
[0088] Embodiment 1: Ultra-compressed sensing channel estimation method based on time domain
[0089] In wireless communication systems, time domain channel estimation is an important step to ensure data transmission reliability. Traditional time domain channel estimation methods perform poorly under low signal-to-noise ratio conditions and have high computational complexity. In order to improve the performance of time domain channel estimation, this embodiment uses ultra-compressed sensing technology combined with deep learning methods to estimate the time domain channel.
[0090] Specific implementation steps:
[0091] 1) Channel sampling:
[0092] The sampling subnetwork is used to compress and sample the input time-domain noisy channel image to generate a compressed sensing measurement value. Specifically, the input channel image is divided into several non-overlapping image blocks, and the convolutional layer is used to compress and sample each block. The sampling result is a compressed sensing measurement value tensor.
[0093] 2) Initial reconstruction:
[0094] The preliminary reconstruction subnetwork receives the compressed sensing measurement value, performs preliminary reconstruction of the channel image through convolution operation, and generates a preliminary reconstructed channel image. This preliminary reconstruction process mainly simulates the linear mapping from the compressed sensing measurement value to the channel image, ensuring the basic accuracy of the preliminary channel estimation.
[0095] 3) Deep reconstruction:
[0096] The initially reconstructed channel image is input into the deep reconstruction subnetwork and further processed by a deep convolutional residual network with 15 convolutional layers to generate the final time domain channel estimation result. Through the design of residual blocks and dilated convolutions, the network can capture a wider range of time domain channel features, thereby improving the estimation accuracy.
[0097] In this embodiment, by adopting ultra-compressed sensing technology, the amount of sampling data required for time domain channel estimation is greatly reduced. At the same time, combined with deep learning methods, the accuracy of channel estimation is significantly improved, especially in low signal-to-noise ratio environments.
[0098] Embodiment 2: Frequency-domain based ultra-compressed sensing channel estimation method
[0099] Frequency domain channel estimation is crucial in multi-carrier modulation (such as OFDM) systems. Traditional frequency domain channel estimation methods face the challenges of high computational complexity and estimation errors. To overcome these problems, this embodiment proposes a frequency domain based ultra-compressed sensing channel estimation method.
[0100] 1) Channel sampling:
[0101] The input frequency domain channel image is compressed and sampled through the sampling subnetwork. The input frequency domain channel image is also divided into several non-overlapping image blocks, and the convolutional layer samples each block to obtain the compressed sensing measurement value of the frequency domain channel.
[0102] 2) Initial reconstruction:
[0103] The compressed sensing measurements are initially reconstructed using the preliminary reconstruction subnetwork. The network recovers the preliminary frequency domain channel image from the compressed sensing measurements by simulating linear mapping, and outputs the preliminary reconstructed frequency domain channel image.
[0104] 3) Deep reconstruction:
[0105] The initially reconstructed frequency domain channel image is input into the deep reconstruction sub-network, and finely reconstructed through the deep convolutional residual network to generate the final frequency domain channel estimation result. Through the feature extraction of the convolution layer and the fine processing of the residual block, the deep reconstruction sub-network can capture the complex characteristics of the frequency domain channel and improve the accuracy of channel estimation.
[0106] This embodiment reduces the computational complexity of frequency domain channel estimation through ultra-compressed sensing technology, and at the same time enhances the accuracy of channel estimation by using deep learning methods, especially in scenarios with severe frequency selective fading, which can significantly improve the reliability and accuracy of channel estimation.
[0107] like Figure 3 As shown, the channel estimation method provided by the embodiment of the present invention includes:
[0108] S1, the sampling subnetwork compresses and samples the input noisy channel image to obtain the compressed sensing measurement value;
[0109] S2, the two reconstruction sub-networks try to learn an end-to-end mapping from channel measurements to the true channel;
[0110] S3, the preliminary reconstruction sub-network performs a linear preliminary reconstruction on the channel measurement values;
[0111] S4, then performs a non-linear deep reconstruction through the deep reconstruction sub-network.
[0112] For channel estimation based on deep denoising, its network architecture is derived from image processing, where the signal-to-noise ratio of images is highly fixed and the dynamic range of change is small. In wireless communication, however, the dynamic range of change in the signal-to-noise ratio brought about by user mobility is very large, which makes the applicability of the deep denoising network in wireless communication systems poor and its performance suffer losses. In order to effectively handle mobile communication scenarios with a large dynamic range of signal-to-noise ratio using only one network, the present invention designs a new convolutional neural network model from the perspective of image reconstruction.
[0113] Specifically, the present invention proposes a deep neural network based on compressive sensing for estimating channels with a large dynamic range of signal-to-noise ratio. The compressive sensing problem can be described as: given a signal Based on its measurement value Reconstruct the original signal x,
[0114] y = Φx + n (7)
[0115] where the matrix is the observation matrix or sampling matrix, is the observation noise. Since m < n, this system of equations is underdetermined and cannot be directly solved. The compressive sensing theory proves the solvability under the condition that x is sparse. Assuming that x is acted on by a transformation matrix (also known as the sparse basis matrix), then the solution of equation (7) can be transformed into the following optimization problem:
[0116]
[0117] where ∈ is the bound of the noise value. The main challenges in solving the compressive sensing problem are designing the sampling matrix and the reconstruction algorithm.
[0118] Existing model-based iterative algorithms have high requirements on the properties of the sampling matrix and the prior knowledge of the signal. Generally, the sparse domain, i.e., the matrix Ψ, needs to be manually given. The sampling matrix Φ and the sparse matrix Ψ need to satisfy low correlation, and the sampling matrix Φ and the sparse matrix Ψ need to satisfy the constraint equidistance condition. To meet these requirements, random Gaussian or Bernoulli matrices are generally selected as sampling matrices. These sampling matrices are independent of the signal and cannot adaptively extract specific structural features in the real-world signal, resulting in poor estimation accuracy. In addition, these iterative algorithms have large computational complexity and are difficult to meet real-time requirements. Neural networks can adaptively optimize the sampling and reconstruction process based on training samples, and obtain higher reconstruction accuracy by extracting and learning data features. In addition, using neural networks to replace traditional sampling matrices and iterative reconstruction models greatly relaxes the prior requirements on the properties of the sampling matrix and the sparsity of the signal, and does not require the sparse domain where the signal is located to be specified in advance. Moreover, the characteristics of separation of complexity and performance of neural networks make the delay of online reasoning much shorter than that of traditional iterative reconstruction methods.
[0119] Taking into account the requirements for versatility and accuracy of channel estimation, the present invention proposes the following Figure 4 The data-driven deep compression reconstruction neural network CSCENet is shown. The entire network is divided into three sub-networks from left to right: a sampling sub-network, a preliminary reconstruction sub-network, and a deep reconstruction sub-network. The sampling sub-network compresses and samples the input noisy channel image to obtain a compressed sensing measurement value. The subsequent two reconstruction sub-networks attempt to learn an end-to-end mapping from the channel measurement value to the true channel, where the preliminary reconstruction sub-network performs a linear preliminary reconstruction of the channel measurement value, and then performs a nonlinear deep reconstruction through the deep reconstruction sub-network. As far as the present invention is aware, this is the first channel reconstruction network based on deep compressed sensing that is fully implemented using a data-driven approach.
[0120] Since general neural networks are real-valued, the complex-valued channel image Where M×K corresponds to the dimensions of the channel in the spatial domain and the time domain respectively. The real and imaginary parts need to be separated before inputting into the network. In the current research on channel estimation, there are several different ways to process the real and imaginary parts of the separated channel image. The first is to treat the real and imaginary parts as two independent images and input them into the network separately; the second is to stack the real and imaginary parts left and right into a real matrix with twice the number of columns. Then input the network; the third is to regard the real part and the imaginary part as two channels of the image, that is, to connect them in a new dimension Generally speaking, the real and imaginary parts of the channel matrix can be considered orthogonal. The first two processing methods cannot well extract the channel phase information implicit in the complex number. Therefore, the present invention adopts the third method, that is, the size of the network input data is M×K×2. The size of the network output is the same as the input, and finally it is merged into complex data.
[0121] Sampling subnetwork:
[0122] In order to reduce the complexity of sampling, the compressed sampling subnetwork adopts block-wise compressed sampling, which divides the input two-channel image into (M / B)×(K / B) non-overlapping image blocks with a block size of B×B×2. B can be adjusted and needs to be divisible by both M and K. Then, block-wise sampling is implemented through a convolutional layer, where n is used. b A B×B×2 convolution filter, the stride of the convolution kernel is set to B×B so that it acts on different blocks when sliding, and no zero padding is used. b Represents the number of compressed sensing measurements obtained for each block. When the sampling rate is set to δ, n b =δ·2B 2 In fact, this convolutional layer simulates the sampling process of traditional block compressed sensing, where each block uses the same sampling matrix:
[0123] y b =Φ b x b (9)
[0124] in is the vectorized image block, is the block sampling matrix, is the compressed sensing measurement value of the image block.
[0125] Preliminary reconstruction of the sub-network:
[0126] The purpose of the preliminary reconstruction network is to preliminarily reconstruct each image block from the compressed sensing measurements through a convolutional layer. After passing through the sampling subnetwork, all the output measurements form a (M / B)×(K / B)×n b A tensor where the measurement value of each block is a 1×1×n b The present invention uses 2B 2 1×1×n b The convolution filter is used to reconstruct all blocks, and the convolution kernel step size is 1×1. Then, the size of the output tensor is changed from (M / B)×(K / B)×2B through the "Reshape+concatenate" function. 2 Reshape to M×K×2 to obtain the preliminary reconstructed channel image. The convolution layer of the preliminary reconstruction network simulates the following matrix operation:
[0127]
[0128] in is the initial reconstructed image block after vectorization, is the initial reconstruction matrix. Φb and The first two sub-networks simulate linear operations, so no activation function is used.
[0129] Deep reconstruction sub-network:
[0130] The purpose of the deep reconstruction subnetwork is to further refine the channel image and improve the reconstruction quality. The deep reconstruction subnetwork contains a total of 15 convolutional layers. First, the initially reconstructed M×K×2 channel image is input into the first convolutional layer for feature extraction to generate 64 feature maps, which then pass through 6 cascaded residual blocks and finally pass through two ordinary convolutional layers (the first of which uses the ReLU function) to form the final M×K×2 reconstructed channel image. The residual block contains two batch normalization (BN) layers, a ReLU function and a dilated convolution layer, as well as a short skip connection, such as Figure 3 As shown in the figure. The skip connection used in the residual block promotes gradient propagation without losing local information, thus reducing the difficulty of training. The dilation rate used in the dilated convolution is 4. The traditional convolution operation only considers the adjacent input pixels at each position, while the dilated convolution introduces holes in the convolution kernel, which can expand the receptive field while keeping the number of parameters unchanged, so that the network can utilize a wider range of spatial information. In the deep reconstruction subnetwork, the spatial size of the convolution kernel is 3×3, and the number of feature channels in the middle layer is 64.
[0131] The channel estimation network CSCENet based on deep compressed sensing proposed in the present invention can achieve robust performance only in the time-frequency domain, with a large SNR dynamic range and in scenarios where users move, avoiding additional hardware (3D lens antenna array) or software (beam domain spatial transformation) costs.
[0132] The present invention establishes a channel model for simulation according to 3GPP TR 38.901, and the frequency band used is 5 GHz. Figure 5 and Figure 6 As shown in the figure, CSCENet performs well in a large dynamic SNR range [-20dB 10dB]. At a SNR of -15dB, CSCENet achieves nearly 21dB and 7dB gains compared to LS and LMMSE, respectively. For high SNR scenarios, CSCENet can achieve almost the same performance as LMMSE (SNR from 5 to 10dB).
[0133] In addition to the high online computational complexity, LMMSE also requires prior statistical information of the channel, which often requires considerable resources to obtain in practice. Therefore, LMMSE is usually only used as a benchmark for measuring the quality of channel estimation in theory, but is not feasible in practice. However, the complexity and performance of CSCENet are separated. The most complex part is completed by offline training, and the weight coefficients are obtained and applied to the online reasoning stage. In this way, online reasoning is just a simple addition, subtraction, multiplication and division operation. The online calculation complexity is much smaller than LMMSE, and there is no need to obtain prior knowledge of the channel.
[0134] It is worth noting that after beam transformation, the channel will be sparser and CSCENet will achieve better performance.
[0135] An application embodiment of the present invention provides a computer device, which includes a memory and a processor. The memory stores a computer program. When the computer program is executed by the processor, the processor executes the steps of the channel estimation method.
[0136] An application embodiment of the present invention provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, the processor executes the steps of the channel estimation method.
[0137] An application embodiment of the present invention provides an information data processing terminal, which includes a channel estimation system.
[0138] The CSCENet proposed in the present invention is suitable for fields that require channel estimation, including but not limited to: communication equipment (core network, base station, terminal, etc.) of 3GPP standard protocol, WiFi equipment of IEEE standard protocol, Bluetooth equipment, transceiver equipment of Zigbee standard protocol, transceiver equipment of Lora protocol, and transceiver equipment of wireless Mesh protocol. The traditional channel estimation algorithm starts from the underlying logic of energy and assumes that the energy of noise and interference is less than the useful signal. This has two disadvantages. One is that the performance is greatly reduced in low signal-to-noise ratio scenarios. The second is the coupling of its complexity and performance. From formula (4), the traditional theoretical optimal LMMSE channel estimation algorithm is based on the channel autocorrelation matrix R hh =E{hh H} is based on known information. In practice, the channel needs to be estimated. In order to ensure that the channel autocorrelation matrix required by LMMSE reaches sufficient accuracy, the channel needs to be estimated and counted for a long time, which is often not feasible in practice. At present, the traditional LS or LS-like channel estimation is widely used in the actual industry. Although it does not require long-term channel statistical information, its channel estimation performance is greatly reduced, especially in low signal-to-noise ratio scenarios. The deep learning-based CSCENet proposed in the present invention starts from the underlying logic of feature recognition and can identify and capture the features of channel signals. In wireless communication systems, there are obvious differences between channel signals, noise and interference, which makes the CSCENet proposed in the present invention show a more robust performance on the impact of signal energy size, especially at low signal-to-noise ratios. It still has good channel feature recognition and channel estimation capabilities.
[0139] It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware part can be implemented using dedicated logic; the software part can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated design hardware. It can be understood by a person of ordinary skill in the art that the above-mentioned devices and methods can be implemented using computer executable instructions and / or contained in a processor control code, such as a carrier medium such as a disk, CD or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. Such code is provided on the carrier medium. The device and its modules of the present invention can be implemented by hardware circuits such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, etc., or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., can also be implemented by software executed by various types of processors, and can also be implemented by a combination of the above-mentioned hardware circuits and software, such as firmware.
[0140] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with the technical field within the technical scope disclosed by the present invention and within the spirit and principle of the present invention should be covered by the protection scope of the present invention.
Claims
1. A channel estimation method, characterized in that: include: S1, the sampling subnetwork compresses and samples the input noisy channel image to obtain the compressed sensing measurement value; S2, the two reconstruction sub-networks try to learn an end-to-end mapping from channel measurements to the true channel; S3, the preliminary reconstruction subnetwork performs a linear preliminary reconstruction of the channel measurement values; S4 then performs nonlinear deep reconstruction through the deep reconstruction sub-network.
2. The channel estimation method according to claim 1, characterized in that: The convolution layer is used to simulate and replace the compressed sensing sampling, and the compressed sensing measurement value is obtained as the initial feature map of the channel; The mapping relationship between the compression sense measurement value and the real channel is obtained through convolution layer training, and the weight coefficient of the mapping relationship is obtained; The weight coefficient of the mapping relationship between the trained compressed sensing measurement value and the real channel is used for online channel estimation, and a preliminary reconstructed channel is obtained according to the input compressed sensing measurement value; The weight coefficients of the mapping relationship between the preliminary reconstructed channel and the real channel are obtained through deep convolutional residual network training; The weight coefficient of the mapping relationship between the trained preliminary reconstructed channel and the real channel is used for online channel estimation, and the deep reconstructed channel is obtained according to the input preliminary reconstructed channel; The new compressed sensing channel estimation network can estimate the wireless channel in the following domains: time domain, frequency domain, antenna domain, and beam domain.
3. The channel estimation method according to claim 1, characterized in that: The sampling subnetwork specifically includes: In order to reduce the complexity of sampling, the compressed sampling subnetwork adopts block-by-block compressed sampling, which divides the input two-channel image into (M / B)×(K / B) non-overlapping image blocks with a block size of B×B×2. B can be adjusted and needs to be divisible by both M and K. Then, a convolutional layer is used to implement block-by-block sampling, where n is used. b A B×B×2 convolution filter, the convolution kernel step size is set to B×B so that it acts on different blocks when sliding, and no zero padding is used; n b Represents the number of compressed sensing measurements obtained for each block. When the sampling rate is set to δ, n b =δ·2B 2 ; In fact, this convolutional layer simulates the sampling process of traditional block compressed sensing, where each block uses the same sampling matrix: y b =Φ b x b (9) in is the vectorized image block, is the block sampling matrix, is the compressed sensing measurement value of the image block.
4. The channel estimation method according to claim 1, characterized in that: The initial reconstruction sub-network specifically includes: The purpose of the preliminary reconstruction network is to reconstruct each image block from the compressed sensing measurements through a convolutional layer; after passing through the sampling subnetwork, all the output measurements form a (M / B)×(K / B)×n b A tensor where the measurement value of each block is a 1×1×n b Vector; use 2B 2 1×1×n b The convolution filter is used to reconstruct all blocks, and the convolution kernel step size is 1×1. Then, the size of the output tensor is changed from (M / B)×(K / B)×2V through the "Reshape+concatenate" function. 2 Reshape to M×K×2 to obtain the preliminary reconstructed channel image; the convolution layer of the preliminary reconstruction network simulates the following matrix operation: in is the initial reconstructed image block after vectorization, is the initial reconstruction matrix, Φ b and are all obtained through learning; the first two sub-networks simulate linear operations, so no activation function is used.
5. The channel estimation method according to claim 1, wherein: The deep reconstruction sub-network specifically includes: The purpose of the deep reconstruction subnetwork is to further refine the channel image and improve the reconstruction quality. The deep reconstruction subnetwork contains a total of 15 convolutional layers. First, the initially reconstructed M×K×2 channel image is input into the first convolutional layer for feature extraction to generate 64 feature maps, and then passes through 6 cascaded residual blocks and finally two ordinary convolutional layers to form the final M×K×2 reconstructed channel image. The residual block contains two batch normalization layers, ReLU function and dilated convolution layer, as well as a short skip connection. The skip connection used in the residual block promotes gradient propagation without losing local information, thus reducing the difficulty of training. The dilation rate used in the dilated convolution is 4. The traditional convolution operation only considers the adjacent input pixels at each position, while the dilated convolution introduces holes in the convolution kernel, which can expand the receptive field while keeping the number of parameters unchanged, thereby enabling the network to utilize a wider range of spatial information. In the deep reconstruction subnetwork, the spatial size of the convolution kernel is 3×3, and the number of feature channels in the middle layer is 64.
6. A channel estimation system for implementing the channel estimation method according to any one of claims 1 to 5, characterized in that: include: A sampling subnetwork module is used to compress and sample the input noisy channel image to obtain a compressed sensing measurement value; A preliminary reconstruction subnetwork module is used for performing a linear preliminary reconstruction of the channel measurement values; Deep reconstruction sub-network module, used for nonlinear deep reconstruction.
7. A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the channel estimation method according to any one of claims 1 to 5.
8. A computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes the steps of the channel estimation method according to any one of claims 1 to 5.
9. An information data processing terminal, comprising the channel estimation system according to claim 6.
Citation Information
Cited By
Method and apparatus for channel noise reduction and shaping in a wireless communication system
WO2026160870A1
