A knowledge-driven deep learning method for fast estimation of TSN high-dimensional channel

Abstract

The application relates to a knowledge-driven deep learning TSN high-dimensional channel fast estimation method, and belongs to the technical field of wireless communication channel estimation.The application designs a GMMV-LAMP high-dimensional channel fast estimation network based on knowledge driving, and a frequency-selective wideband redundant dictionary design method in channel estimation.The GMMV-LAMP network can quickly and accurately perform wideband high-dimensional channel estimation.The frequency-selective wideband redundant dictionary can accurately obtain the sparse domain representation of the channel in the millimeter wave UM-MIMO pure far field, near field and far-near field mixed field, and can overcome the beam squint effect caused by the frequency flatness dictionary.The application applies the redundant dictionary to the GMMV-LAMP network, accurately and quickly performs millimeter wave wideband large-scale MIMO channel estimation in the far-near field mixed field, and can exert the advantages of millimeter wave UM-MIMO technology, such as low delay and high bandwidth, for TSN.The application improves the accuracy of UM-MIMO high-dimensional channel estimation, and has a faster convergence speed.

Description

Technical Field

[0001] This invention relates to a knowledge-driven deep learning-based fast estimation method for high-dimensional channels in Time Sensitive Networks (TSN), belonging to the field of wireless communication channel estimation technology. Background Technology

[0002] With the evolution of sixth-generation (6G) mobile communication systems, various new service scenarios have placed higher demands on the real-time performance of Ethernet, and the concept of Time Sensitive Networking (TSN) has gradually attracted widespread attention.

[0003] Time-Sensitive Networking (TSN) aims to provide data transmission capabilities with deterministic latency, thereby ensuring the security and reliability of application systems. Communication physical layer technology plays a crucial role in achieving highly reliable and high-speed communication in TSN.

[0004] Ultra Massive Multi-Input Massive Output (UM-MIMO) is one of the key physical layer technologies in sixth-generation (6G) mobile communication systems, providing sufficiently large array gain. Higher frequency bands such as millimeter wave (mmWave) and terahertz (THz) offer abundant bandwidth spectrum resources and ultra-high data rates, significantly increasing system throughput. Therefore, millimeter-wave UM-MIMO technology can provide greater transmission bandwidth and shorter network latency for time-sensitive networks (TSNs). High-precision channel estimation helps obtain information such as multipath delay, signal strength, noise, and interference. This is crucial for optimizing higher-layer network parameters and ensuring the reliability and efficiency of network data transmission. Low-complexity channel estimation algorithms can also reduce the latency of signal processing and propagation in the physical layer, thus providing the physical layer foundation for deterministic, low-latency communication in the TSN network layer.

[0005] To fully leverage the advantages of millimeter-wave UM-MIMO technology in time-sensitive networks, accurate and rapid channel estimation is essential.

[0006] Due to the massive number of antennas, channel estimation at the access point (AP) in millimeter-wave UM-MIMO systems—that is, obtaining complete downlink high-dimensional channel state information (CSI)—is challenging. First, the large number of UM-MIMO antennas introduces significant pilot and computational overhead. In time-division duplex (TDD) massive MIMO systems using all-digital arrays, the relatively limited number of UE antennas and good channel reciprocity between uplink and downlink channels allow the AP to typically estimate the uplink CSI from the pilots transmitted by the UE to infer the downlink CSI. However, in traditional sub-6 GHz frequency-division duplex (FDD) massive MIMO systems using all-digital arrays, uplink and downlink channel reciprocity is absent, making it impossible to obtain the downlink CSI by estimating the uplink CSI. Furthermore, radio frequency calibration in millimeter-wave / terahertz systems is difficult, and uplink and downlink reciprocity deteriorates in TDD-based millimeter-wave / terahertz systems. Millimeter-wave / terahertz MIMO systems typically employ hybrid analog-to-digital arrays, and the uplink channel pilot overhead is proportional to the number of receiver antennas. In contrast, the downlink channel estimation training time can be relatively small because multiple UEs can simultaneously perform channel estimation (CE) based on the downlink pilot signal broadcast by the AP. Therefore, the AP obtains downlink CSI information by broadcasting pilots and performing downlink channel estimation and CSI feedback at the UEs.

[0007] Due to the angular domain sparsity of MIMO channels, various channel estimation algorithms based on compressed sensing (CS) have been proposed, including greedy algorithms such as Orthogonal Matching Pursuit (OMP) and Bayesian inference methods such as Approximate Message Passing (AMP). Utilizing prior information such as the structured sparsity of the channel can further improve the accuracy of these algorithms. However, limited by manual prior design and iterative methods, traditional algorithms suffer from poor environmental adaptability and high convergence time overhead, leading to a performance degradation in the envisioned 6G communication scenarios.

[0008] In recent years, deep learning (DL) has been considered a key enabling technology in the field of communications. By using deep learning to train a neural network model with a preset objective function, it is possible to learn to adapt to the characteristics of real-world data, and the trained model can make real-time predictions with low complexity.

[0009] Currently, deep learning methods combined with wireless physical layer transmission algorithms include data-driven and knowledge-driven approaches. Data-driven deep learning methods rely on large amounts of labeled sample data to train and generate communication signal processing modules, while knowledge-driven deep learning methods utilize mature algorithm models. Compared to data-driven methods, knowledge-driven deep learning offers better reliability and interpretability. For example, there is a knowledge-driven deep learning-based Multiple Measurement Vector-Learned AMP (MMV-LAMP) algorithm. This algorithm performs a deep expansion of the MMV-AMP algorithm, replacing some parameters with those trained through deep learning. Compared to traditional methods, this method exhibits faster convergence speed and higher channel estimation accuracy.

[0010] Traditional compressed sensing methods typically employ a DFT-based redundant dictionary to transform the spatial domain into the virtual angle domain, assuming that electromagnetic waves propagate according to far-field plane wave characteristics and that CSI (Channel Sequencing Index) is sparsity in the virtual angle domain, thus performing channel estimation. However, in UM-MIMO systems, due to the large antenna array size and increased Rayleigh distance, near-field and far-field mixing occurs in practical applications, no longer satisfying the plane wave propagation assumption. The sparsity of the virtual angle domain deteriorates, and the effectiveness of traditional channel estimation methods utilizing virtual angle domain sparsity decreases.

[0011] Furthermore, in traditional wideband MIMO systems, the incident virtual angles of electromagnetic waves on all subcarriers are approximately the same. When using a frequency flatness dictionary, the virtual angle domain channels on different subcarriers have approximately the same support set, allowing algorithms such as MMV-LAMP to achieve good channel estimation accuracy. However, when the system bandwidth increases dramatically, significant differences in wavelengths on different subcarriers lead to virtual angle shifts, causing changes in the support set positions of the virtual angle domain channels on different subcarriers. In this case, using a frequency flatness dictionary for sparse transformation results in a severe loss of channel estimation accuracy. Summary of the Invention

[0012] The purpose of this invention is to address the technical problems of decreased channel estimation accuracy and slow convergence speed in existing channel estimation techniques for ultra-large-scale millimeter-wave broadband MIMO systems in mixed far-field and near-field scenarios. This invention creatively proposes a knowledge-driven deep learning-based fast channel estimation method for high-dimensional TSN.

[0013] This invention includes the design of a knowledge-driven GMMV-LAMP (Generalized MMV-LAMP) high-dimensional channel fast estimation network, and a frequency-selective broadband redundancy dictionary design method for channel estimation.

[0014] Among them, the GMMV-LAMP network can perform broadband high-dimensional channel estimation quickly and accurately. The frequency-selective broadband redundant dictionary can accurately obtain the sparse domain representation of the channel in pure far-field, near-field, and mixed far-near-field scenarios of millimeter-wave UM-MIMO, and overcome the beam-squinting effect caused by the frequency-flatness dictionary. In this invention, by applying the redundant dictionary to the GMMV-LAMP network, accurate and rapid millimeter-wave broadband large-scale MIMO channel estimation is performed in mixed far-field and near-field scenarios, thereby leveraging the advantages of millimeter-wave UM-MIMO technology for TSN, such as low latency and high bandwidth.

[0015] The objective of this invention is achieved through the following technical solution.

[0016] First, this invention designs a knowledge-driven GMMV-LAMP channel fast estimation network. By deeply expanding the traditional AMP algorithm and introducing trainable parameters and fully connected layers, the GMMV-LAMP channel fast estimation network is obtained.

[0017] Specifically, the channel estimation problem is first modeled as a multi-vector sparse recovery problem. Each observation involves K dimensions. The single-vector observation is extended to multi-vector observation (MMV), and for different observation dimensions k, 1 ≤ k ≤ K, the corresponding observation matrix A[k] is used. The observation dimension k corresponds to different subcarrier frequencies in the channel estimation problem of this invention. This is then transformed into a generalized multi-vector (GMMV) sparse recovery problem, namely:

[0018] Y[k]=A[k]H sparse [k]+N

[0019] Where Y[k] is the noisy observation signal of the k-th subcarrier, with M rows and 1 column; A[k] is the observation matrix of the k-th subcarrier, with M rows and V columns; H sparse [k] is the sparse domain channel matrix on the k-th subcarrier, with V rows and 1 column, and H sparse It has a common support set sparsity on different subcarriers, M << V; N represents complex Gaussian white noise.

[0020] This invention addresses the aforementioned GMMV sparsity recovery problem using a GMMV-LAMP network. The GMMV-LAMP network is specifically designed as follows:

[0021] In the AMP algorithm, the fixed observation matrix is ​​replaced with a trainable matrix B[k], which determines the prior parameters θ. t The contraction function η(·; θ) t ,) is replaced with an MMSE denoiser η with a trainable prior non-zero probability γ and a trainable variance ε. CS (·;γ,ε,·).

[0022] The AMP algorithm is expanded in depth for each iteration, different observation matrices are used for different subcarriers, and several fully connected layers are introduced to obtain the GMMV-LAMP network. The network structure is as follows: Figure 2 As shown. The network takes the received signal Y as input and outputs an estimated antenna-frequency domain channel matrix.

[0023] Furthermore, the specific computational process of the GMMV-LAMP network is as follows:

[0024] First, the matrix is ​​initialized by initializing the initial residual V0 with the received signal Y, and the initial sparse estimate H. sparse，0 The zero matrix is ​​used, and the trainable observation matrix B[k] for each subcarrier is initialized to A[k].

[0025] Next, the network is solved iteratively. The network has TN layers, and the forward propagation steps of the t-th layer are as follows:

[0026] Step 1: Update the noisy estimate of the sparse matrix of the t-th layer network. For each subcarrier k, calculate the noisy estimate of the sparse matrix of the t-th layer network.

[0027] Step 2: Update the estimated value of the sparse matrix of the t-th layer network. Calculate the input residual v of this layer network in the receiving time slot. t-1 The average value of [k] Compared with the noisy estimate obtained in the previous step Input contraction function η CS In this study, the sparsity of the common support set is utilized to calculate the estimated value of the sparse matrix of the t-th layer network.

[0028] Step 3: Update the residual value v of the t-th network. t [k]. Based on the received signal Y[k] of the k-th subcarrier, the estimated value of the sparse matrix of the t-th layer network. and the residual v of the (t-1)th layer network t-1 [k], calculate the residual v of this layer. t [k].

[0029] Step 4: Estimate the sparse domain channel matrix of the t-th layer network. Transformation yields the antenna-frequency domain channel matrix Complete the forward propagation of this layer. Calculate the NMSE (Normalized Mean Square Error) loss function, and perform backpropagation and parameter updates. After iterative convergence of the TN layer network, output the estimated antenna-frequency domain channel matrix.

[0030] On the other hand, in order to solve the technical problem of decreased channel sparsity in the near-field virtual angle domain of UM-MIMO, this invention proposes a frequency-selective broadband redundancy dictionary design method.

[0031] In the GMMV-LAMP network, the broadband redundancy dictionary used in pure far-field, mixed far-near-field, and near-field scenarios is generated using the following method:

[0032] For the pure far-field case, a DFT redundancy dictionary is used on each subcarrier, and the DFT redundancy dictionaries on each subcarrier together constitute a broadband redundancy dictionary.

[0033] For cases involving a mixture of near-field and far-near-field conditions, the broadband redundancy dictionary is generated using a data-driven deep learning method, with the following steps:

[0034] Step 1: Within the given Rayleigh distance of the UM-MIMO antenna array, in polar coordinates... In the form of , V points are randomly generated, where V is the dimension of the generated sparse domain vector, and d represents the distance in polar coordinates. This represents the azimuth angle in polar coordinates.

[0035] Step 2: Substitute the initial coordinates into the spherical waveguide vector to calculate the initial values ​​of the data-driven broadband redundancy dictionary on each subcarrier. Use the coordinates of V points as trainable parameters, and apply the generated broadband redundancy dictionary to the GMMV-LAMP network, optimizing and updating it along with the training of the GMMV-LAMP network.

[0036] Step 3: Through iterative updates of multiple layers of network, the generated redundant dictionary D is obtained. Learn .

[0037] A data-driven broadband redundancy dictionary is applied to channel estimation in a GMMV-LAMP network. The execution process is as follows: Figure 3 As shown.

[0038] Beneficial effects

[0039] This invention overcomes the beam squint effect in channel estimation of broadband millimeter-wave UM-MIMO systems in existing technologies, solves the problem of virtual angle domain sparsity loss in millimeter-wave UM-MIMO systems under near-field and far-near-field mixed conditions, and improves the accuracy of high-dimensional channel estimation for UM-MIMO. Compared with existing methods, this invention has a faster convergence speed. Attached Figure Description

[0040] Figure 1 This is a flowchart of the present invention;

[0041] Figure 2 This is a structural diagram of the GMMV-LAMP channel estimation network proposed in this invention;

[0042] Figure 3 The following is a flowchart illustrating the execution of the data-driven broadband redundancy dictionary proposed in this invention when applied to the GMMV-LAMP channel estimation network.

[0043] Figure 4 The graph shows a comparison of the convergence speed of existing methods and the GMMV-LAMP network proposed in this invention in pure far-field scenarios.

[0044] Figure 5 This is a comparison chart of the channel estimation accuracy performance between the pure near-field scenario and the existing redundant dictionary scheme.

[0045] Figure 6 The figure shows a comparison of the accuracy of the channel estimation method proposed in this invention in mixed near-field scenarios with different numbers of far-field scatterers. Detailed Implementation

[0046] The method of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0047] Example

[0048] This invention proposes a knowledge-driven deep learning-based fast estimation method for high-dimensional channels in broadband millimeter-wave UM-MIMO systems, targeting far-field, near-field, and mixed far-near-field scenarios in TSN, for high-dimensional channel estimation of UM-MIMO systems.

[0049] The technical solution of the present invention will be illustrated by taking a typical single-cell multi-user millimeter-wave broadband UM-MIMO system as an example.

[0050] In a typical single-cell multi-user millimeter-wave broadband UM-MIMO system model, the AP adopts a hybrid analog-digital MIMO architecture, which has N RF One radio frequency chain and N APEach UE is equipped with a single antenna, using a uniform linear array (ULA). All UEs have the same number of downlink channel scatterers, all of which are L. Communication is achieved using OFDM modulation with K subcarriers. During the channel estimation phase, the AP broadcasts pilot signals to the UEs, with each pilot signal containing G pilot time slots.

[0051] For each UE, the received pilot signal y on the k-th subcarrier DL [k] is represented as:

[0052] y DL [k]=S[k]h DL [k]+n DL [k]

[0053] in, This represents the pilot signals transmitted by the AP in G time slots, where S[k] represents the set of complex numbers, which is known to the UE. This represents the noisy received signal of the UE on the k-th subcarrier. This is the noise introduced by the k-th subcarrier during channel propagation.

[0054] The far-field channel model is based on the plane wave approximation. When the distance is much greater than the Rayleigh distance, the channel vector h on the k-th subcarrier... DL [k] is considered as a function relating only to the AOD of each scatterer, exhibiting sparsity in the virtual angular domain. Let the starting angle of the m-th far-field incident path be... Antenna spacing is λ c λ is the center carrier wavelength. k Given the subcarrier wavelength, the array far-field steering vector is:

[0055]

[0056] Where T represents matrix transpose, e represents natural constant, and j represents imaginary unit.

[0057] The near-field channel model is based on the spherical wave propagation assumption and the millimeter-wave Saleh-Valenzuela channel model. The channel vector on the k-th subcarrier is...

[0058]

[0059] in, β l Let τ be the channel gain of the l-th path. l f represents the time delay relative to the reference antenna on the l-th path; s For system bandwidth; (x l ,yl ) represents the Cartesian coordinates of the l-th scatterer. The spherical wave array steering vector for the k-th subcarrier on the l-th path:

[0060]

[0061] Among them, D i,l =d i,l -d 1,l , It defines the distance between the UE and the i-th antenna of the AP on the l-th path, D. i,l This represents the relative distance between the i-th antenna of the AP and the UE along the l-th path, and between the reference antenna and the UE. Cartesian coordinates (x... l ,y l It can be represented in polar coordinates. d l Represents the distance coordinates of the l-th scatterer. The angle coordinates of the l-th scatterer are shown, indicating that the UM-MIMO near-field channel is related to both angle and distance.

[0062] In millimeter-wave UM-MIMO systems, scatterers may be located in both the near-field and far-field regions of the access point (AP). This is referred to as the mixed near-field and far-field effect in channel modeling, where the far-field assumption no longer holds. In this case, the channel vector on the k-th subcarrier is represented as:

[0063]

[0064] Among them, L n and L f The number of scatterers in the near and far fields represents the number of scatterers, and L = L n +L f .

[0065] Since the position of the near-field scatterer is determined by distance and angle, the virtual angular domain sparsity of the channel matrix will decrease in near-field and mixed near-field / far-field scenarios of millimeter-wave UM-MIMO. Furthermore, in narrowband systems, all subcarriers have almost the same wavelength, and the antenna spacing is designed to be half the wavelength of the center subcarrier. Therefore, for both near-field and far-field cases, the array steering vector of all subcarriers is considered to be the same, i.e. and For all subcarriers k, k is the same. However, for ultra-large broadband systems with bandwidth comparable to carrier frequency, the array steering vector is frequency-selective in both the near-field and far-field cases; this phenomenon is commonly referred to as beamsight. When the same DFT dictionary is used directly on all subcarriers in a broadband UM-MIMO system, the virtual angular domain support set will shift across different subcarriers due to beamsight, degrading the performance of the GMMV-LAMP channel estimation algorithm.

[0066] To address the aforementioned problems, this invention proposes a method employing a broadband redundant dictionary, which effectively solves the issues of decreased sparsity in the virtual angle domain and beam squint. Specifically:

[0067] Step 1: Calculate the broadband redundancy dictionary in the GMMV-LAMP network for the pure far-field, mixed far-near-field, and near-field scenarios, respectively, based on the different scenarios where the UE is located: far-field, mixed far-near-field, and near-field.

[0068] Step 1.1: Generate the DFT wideband redundancy dictionary.

[0069] When the scatterer is located in the pure far-field region of the UE, the virtual angle domain exhibits sparsity. In this case, the DFT redundancy dictionary is used as the broadband redundancy dictionary in the GMMV-LAMP network, and the DFT redundancy dictionary D on the k-th subcarrier... AD The calculation method for [ρ,k] is as follows:

[0070]

[0071] Where ρ represents the set redundancy factor; That is, the array far-field steering vector In Replace with

[0072] Step 1.2: Generate initial values ​​for the data-driven broadband redundancy dictionary.

[0073] When the scatterer is located in the near field or a mixed near-far field region of the UE, the sparsity of the virtual angle domain of the channel matrix decreases. According to the principle of polar coordinate codebook, the sparsity of the near-field UM-MIMO channel transform domain mainly comes from reasonable sampling points. Therefore, a data-driven broadband redundant dictionary is used to adaptively learn the sparsity characteristics of the near-field scattering propagation environment.

[0074] Step 1.2.1: Randomly generate the polar coordinates of V points. As initial values, the distance and angle components of the above V points are combined into a trainable vector c. d =[d0,d1,…,d V-1 ] T and The initial values ​​of the data-driven broadband redundancy dictionary are calculated. The data-driven broadband redundancy dictionary D learn The vth column is the near-field steering vector corresponding to the Cartesian coordinates of the vth point in the trainable vector.

[0075] Step 1.2.2: Using the aforementioned redundant dictionary as the initial sparse domain transformation dictionary, calculate the observation matrix in the GMMV-LAMP network:

[0076] A[k]=S[k]D Learn [k]

[0077] Where S[k] is known to the UE. Therefore, the training of the data-driven broadband redundancy dictionary can be completed during the GMMV-LAMP channel estimation network iteration process. After the iteration converges, D is calculated from the observation matrix and S[k]. Learn [k].

[0078] Step 2: Channel estimation is performed using the GMMV-LAMP network and a frequency-selective broadband redundancy dictionary.

[0079] By employing a frequency-selective broadband redundancy dictionary, the frequency-antenna domain channel matrix is ​​transformed into a frequency-sparse domain channel matrix with structured sparsity, thus transforming the UM-MIMO channel estimation problem into a GMMV sparse recovery problem. In this example, channel estimation is performed using a T-layer GMMV-LAMP network based on the proposed Learned Approximate Message Passing (LAMP) algorithm, as shown in the following structure. Figure 2 As shown, the observation matrix A[k] = S[k]D[k] in the network changes with the subcarrier index k. The frequency-selective redundancy dictionary D[k] can be either a DFT broadband redundancy dictionary or a data-driven broadband redundancy dictionary, depending on whether it's a far-field or near-field scenario. When using the DFT redundancy dictionary, the observation matrix A[k] is not updated. However, when using the data-driven broadband redundancy dictionary, the observation matrix A is updated along with the network iterations, and the redundancy dictionary is generated accordingly.

[0080] In this embodiment, the proposed GMMV-LAMP channel estimation network is implemented using the open-source deep learning library PyTorch, and Adam is used as the optimizer.

[0081] In this embodiment, the GMMV-LAMP network parameters are divided into global training parameters (such as ε, γ) and hierarchical training parameters (such as B). t[k]). The network training adopts a layer-by-layer training method. When the training of the t-th layer is completed, only the training parameters of the t-th layer are completed, while the global training parameters are continuously trained in all layers. Among them, the data-driven codebook D... Learn [k] Training is only performed on layers 1 and 2. Once layer 2 is trained, the trainable parameter c is obtained. d , This will no longer change. The GMMV-LAMP network uses the estimated channel... With real channel The normalized mean square error (NMSE) between the two sides is used as the cost function Loss, i.e.:

[0082]

[0083] The GMMV-LAMP network in this embodiment has TN layers, and the calculations for each layer are the same. Therefore, only the t-th layer is described here.

[0084] Step 2.1: Initialization.

[0085] The three inputs to the t-th layer network are the sparse domain channel matrix estimated by the (t-1)-th layer. Updated residual V of the previous layer output t-1 Observation signal Y DL It is the same for each layer.

[0086] In the first layer network, initialize the channel estimation matrix. Initialize residual term V0 = Y DL .

[0087] Step 2.2: Update the channel estimate

[0088] Step 2.2.1: Calculate the noisy observations of the channel matrix corresponding to each subcarrier k.

[0089] The observation matrix A in the t-th layer AMP algorithm t [k] is replaced with a trainable matrix B t [k], to accelerate convergence, B at the start of training for each layer. t [k] is initialized to A t [k]. The calculation is as follows:

[0090]

[0091] Where H represents the matrix conjugate transpose operation.

[0092] Step 2.2.2: Calculate the variance of the residual terms for each subcarrier. Combined into a residual variance matrix ∑ t :

[0093]

[0094]

[0095] Here, diag() represents a diagonal matrix with the elements in parentheses as diagonal elements.

[0096] Step 2.2.3: Update the channel estimate

[0097] Collect noisy observations of each subcarrier channel matrix The v-th element is used to form a new row vector of the noisy channel matrix. The model is as follows:

[0098]

[0099] in, This represents the v-th row vector of a noise-free sparse channel. This is the element in row v and column k of a noise-free sparse channel.

[0100] n t,v The noise term follows a standard normal distribution.

[0101] In order to China Resumption This invention designs a denoiser based on Minimum Mean Square Error (MMSE). Due to the identical support set property of the sparse matrix of the channel, for any subcarrier index k, They all have the same sparsity. Assume the noise-free sparse channel vector to be recovered follows a Bernoulli-Gaussian distribution, then... The prior distribution is modeled as follows:

[0102]

[0103] in, It follows a complex Gaussian distribution, and γ is The probability, ω is hour The variances of these parameters are all trainable parameters.

[0104] Shrinkage function based on MMSE denoising Represented as:

[0105]

[0106]

[0107]

[0108] in, Let v be the row vector of the denoised sparse channel estimation matrix. Let φ(x) be the element in the v-th row and k-th column of the denoised sparse channel estimation matrix, 1≤k≤K, φ(x) be the denoising function, and P be a diagonal matrix of trainable prior parameters.

[0109] when After the calculation is complete, the channel estimate is recombined and updated:

[0110]

[0111] in, The element in the v-th row and k-th column of the denoised sparse channel estimation matrix.

[0112] Step 2.3: Update residual term V t .

[0113] Step 2.3.1: Calculate the gradient term b t .

[0114] Introducing a fully connected layer to update the gradient term b t =[b t [1],…,b t [K] T Among them, b for each subcarrier t [k] is determined by the contraction function relative to The gradient is obtained. The intermediate variable for the gradient of each subcarrier k. After the calculation is complete, they are merged into a vector and passed through a fully connected network g. t () to obtain the denoised b t :

[0115]

[0116]

[0117]

[0118] in, This represents the operation of calculating partial derivatives.

[0119] Step 2.3.2: Calculate the residual term V for this layer. t .

[0120] With the help of the Onsager correction term b for each subcarrier k t [k]v t-1[k] is used to update the residual terms on each subcarrier. After the residual terms of all subcarriers have been calculated, the total residual term V of all subcarriers in the previous layer is used. t-1 Through the fully connected layer f t To introduce a momentum factor using parentheses (), thus accelerating convergence:

[0121]

[0122]

[0123]

[0124] in, Let be the residual intermediate variable of the k-th subcarrier. The total residual intermediate variable for all subcarriers.

[0125] Step 2.4: After iterating through the TN-layer network, output the complete frequency-sparse domain channel estimation matrix of the TN-th layer. The frequency-antenna domain channel estimation matrix is ​​calculated using the frequency-selective redundancy dictionary D[k], 1≤k≤K.

[0126]

[0127]

[0128] This concludes the fast estimation process for UM-MIMO high-dimensional channels in this embodiment.

[0129] The following simulation results Figures 4 to 6 This invention illustrates how the proposed broadband redundant dictionary addresses the channel estimation performance loss caused by beam squint, how the data-driven broadband redundant dictionary is effective in learning the sparsity of near and far fields and mixed fields, and how the knowledge-driven GMMV-LAMP proposed in this invention outperforms traditional schemes in terms of convergence time.

[0130] Figure 4In the pure far-field scenario, the channel estimation accuracy performance of the knowledge-driven GMMV-LAMP network of this invention was compared with that of the existing MMV-LAMP network and GMMV-AMP algorithm under different signal-to-noise ratios, measured by NMSE. Comparing the simulation results curves of "GMMV-LAM (using frequency-selective DFT redundancy dictionary, 5 iterations)," "GMMV-AMP (using frequency-selective DFT redundancy dictionary, 100 iterations)," and "GMMV-AMP (using frequency-selective DFT redundancy dictionary, 80 iterations)," it is evident that the GMMV-LAMP network proposed in this invention achieves better channel estimation accuracy at all signal-to-noise ratios compared to the GMMV-AMP algorithm, with fewer iterations and faster convergence speed. Comparing the curves of "GMMV-LAMP (using frequency-flatness DFT redundancy dictionary, 5 iterations)" and "MMV-LAMP (using frequency-flatness DFT redundancy dictionary, 5 iterations)," it is clear that the GMMV-LAMP network of this invention has better channel estimation accuracy compared to the MMV-LAMP network.

[0131] Furthermore, the channel estimation accuracy was compared between using a frequency-selective broadband redundancy dictionary and a frequency-flat broadband redundancy dictionary in the network. Comparing the curves of "GMMV-LAMP (using a frequency-flat DFT redundancy dictionary, 5 iterations)" and "GMMV-LAMP (using a frequency-selective DFT redundancy dictionary, 5 iterations)," it is evident that when using the same estimation method and number of iterations, the proposed frequency-selective broadband redundancy dictionary significantly outperforms the frequency-non-selective dictionary in channel estimation accuracy. This demonstrates that the frequency-selective broadband redundancy dictionary can address the channel estimation accuracy degradation caused by beam-squinting effects.

[0132] Figure 5 The channel estimation accuracy of the proposed method for GMMV-LAMP networks was compared under pure near-field conditions using existing DFT redundant dictionaries, existing polar coordinate redundant dictionaries, and the data-driven redundant dictionary scheme proposed in this invention. It can be seen that the DFT redundant dictionary has the worst channel estimation accuracy in the near-field condition because the sparsity loss is significant when the near-field scatterers are transformed to the virtual angular domain. Existing polar coordinate codebooks, determined by both distance and angle, are suitable for near-field channel estimation in UM-MIMO, and therefore their performance is superior to that of the DFT codebook. The data-driven frequency-selective redundant dictionary proposed in this invention achieves performance similar to that of the polar coordinate codebook, thus proving the effectiveness of the proposed method.

[0133] Figure 6The robustness of the proposed GMMV-LAMP network in channel estimation accuracy under mixed near-field and far-field conditions is demonstrated when using a data-driven broadband redundant dictionary. With a fixed total number of scatterers L=5, the channel estimation accuracy is obtained for different distributions of far-field and near-field scatterers by varying the number of far-field scatterers. Simulation results show that when the number of far-field scatterers N... far =4, the channel estimation accuracy is optimal; as the number of far-field scatterers decreases, the channel estimation accuracy decreases accordingly, but the change is not significant. When the number of far-field scatterers N is 4, the channel estimation accuracy is optimal. far =1, the accuracy decreased by 2dB compared to the optimal accuracy, but the channel estimation accuracy is still within an acceptable range. The results demonstrate that the data-driven broadband redundancy dictionary proposed in this invention is not only applicable to various mixed near-far field scenarios, but also possesses robustness.

[0134] In summary, the above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A knowledge-driven deep learning-based fast estimation method for high-dimensional channels in TSN, characterized in that, include: Design a knowledge-driven GMMV-LAMP channel fast estimation network, as follows; In the AMP algorithm, the fixed observation matrix is ​​replaced with a trainable matrix B[k], which determines the prior parameters θ. t The contraction function η(·; θ) t ,·) is replaced with an MMSE denoiser η with a trainable prior non-zero probability γ and a trainable variance ε. CS (·;γ,ε,·); The AMP algorithm is expanded in depth for each iteration, different observation matrices are used for different subcarriers, and several fully connected layers are introduced to obtain the GMMV-LAMP network. The network receives the signal Y as input and outputs an estimated antenna-frequency domain channel matrix. The computation process of the GMMV-LAMP network is as follows: First, the matrix is ​​initialized by initializing the initial residual V0 with the received signal Y, and the initial sparse estimate H. sparse,0 The zero matrix is ​​used, and the trainable observation matrix B[k] for each subcarrier is initialized to A[k]. Next, the network is solved iteratively; the network has TN layers, and the forward propagation steps of the t-th layer are as follows: Step 1: Update the noisy estimate of the sparse matrix of the t-th layer network; for each subcarrier k, calculate the noisy estimate of the sparse matrix of the t-th layer network. Step 2: Update the estimated value of the sparse matrix of the t-th layer network; calculate the input residual v of this layer network in the receiving time slot. t-1 The average value of [k] Compared with the noisy estimate obtained in the previous step Input contraction function η CS In this study, the sparsity of the common support set is utilized to calculate the estimated value of the sparse matrix of the t-th layer network. Step 3: Update the residual value v of the t-th network. t [k]; Based on the received signal Y[k] of the k-th subcarrier, the estimated value of the sparse matrix of the t-th layer network. and the residual v of the (t-1)th layer network t-1 [k], calculate the residual v of this layer. t [k]; Step 4: Estimate the sparse domain channel matrix of the t-th layer network. Transformation yields the antenna-frequency domain channel matrix Complete the forward propagation of this layer; calculate the NMSE loss function, and perform backpropagation and parameter updates; After the TN layer network iteratively converges, the estimated antenna-frequency domain channel matrix is ​​output. In the GMMV-LAMP network, a broadband redundancy dictionary method is adopted, that is, the broadband redundancy dictionary used in pure far-field, mixed far-near field, and near-field cases is generated in the following way: For the pure far-field case, a DFT redundancy dictionary is used on each subcarrier, and the DFT redundancy dictionaries on each subcarrier together constitute a broadband redundancy dictionary. For cases involving a mixture of near-field and far-near-field conditions, the broadband redundancy dictionary is generated using a data-driven deep learning method, with the following steps: Step 1: Within the given Rayleigh distance of the UM-MIMO antenna array, in polar coordinates... In the form of , V points are randomly generated, where V is the dimension of the generated sparse domain vector, and d represents the distance in polar coordinates. Represents the azimuth angle in polar coordinates; Step 2: Substitute the initial coordinates into the spherical waveguide vector to calculate the initial value of the data-driven broadband redundancy dictionary on each subcarrier; use the coordinates of V points as trainable parameters, apply the generated broadband redundancy dictionary to the above GMMV-LAMP network, and optimize and update it along with the training of the GMMV-LAMP network. Step 3: Through iterative updates of multiple layers of network, the generated redundant dictionary D is obtained. Learn ; Apply a data-driven broadband redundancy dictionary to channel estimation in a GMMV-LAMP network; Specifically, based on different scenarios such as far-field, mixed far-near-field, and near-field, the broadband redundancy dictionary in the GMMV-LAMP network is calculated for pure far-field, mixed far-near-field, and near-field scenarios, including the following steps: Step 1: Generate a DFT wideband redundancy dictionary; When the scatterer is located in the pure far-field region, the virtual angular domain exhibits sparsity; in this case, the DFT redundancy dictionary is used as the broadband redundancy dictionary in the GMMV-LAMP network, and the DFT redundancy dictionary D on the k-th subcarrier is... AD The calculation method for [ρ,k] is as follows: Where ρ represents the set redundancy factor; That is, the array far-field steering vector In Replace with 0, Step 2: Generate initial values ​​for the data-driven broadband redundancy dictionary; When the scatterer is located in the near field or a mixed near and far field region, the sparsity of the virtual angular domain of the channel matrix decreases. Step 2.1: Randomly generate the polar coordinates of V points. Using 0≤v≤V as the initial value, the distance and angle components of the above V points are combined into a trainable vector c. d =[d0,d1,…,d V-1 ] T and The initial values ​​of the data-driven broadband redundancy dictionary are calculated. Among them, the data-driven broadband redundancy dictionary D learn The vth column is the near-field steering vector corresponding to the Cartesian coordinates of the vth point in the trainable vector; Step 2.2: Using the aforementioned redundant dictionary as the initial sparse domain transformation dictionary, calculate the observation matrix in the GMMV-LAMP network: A[k]=S[k]D Learn [k] Where S[k] is known to the UE; after the iteration converges, D is calculated from the observation matrix and S[k]. Learn [k].

2. The knowledge-driven deep learning-based fast estimation method for high-dimensional TSN channels as described in claim 1, characterized in that, Channel estimation is performed using a GMMV-LAMP network and a frequency-selective broadband redundancy dictionary. The GMMV-LAMP network parameters are divided into global training parameters and hierarchical training parameters; The network training employs a layer-by-layer training approach. When the training of layer t is completed, only the training parameters of layer t are trained, while the global training parameters are continuously trained across all layers. The data-driven codebook D... Learn [k] Training is only performed on layers 1 and 2. Once layer 2 is trained, the trainable parameter c is obtained. d , This will no longer change; the GMMV-LAMP network uses the estimated channel... With real channel The normalized mean squared error between them is used as the cost function Loss, that is: The GMMV-LAMP network has TN layers, and the calculations for each layer are identical. For the t-th layer, we have: Step 1: Initialization; The three inputs to the t-th layer network are the sparse domain channel matrix estimated by the (t-1)-th layer. Updated residual V of the previous layer output t-1 Observation signal Y DL It is the same for each layer; In the first layer network, initialize the channel estimation matrix. Initialize residual term V0 = Y DL ; Step 2: Update the channel estimate Step 2.1: Calculate the noisy observations of the channel matrix corresponding to each subcarrier k. The observation matrix A in the t-th layer AMP algorithm t [k] is replaced with a trainable matrix B t [k], B at the start of training for each layer t [k] is initialized to A t [k]; The calculation is as follows: Where H represents the matrix conjugate transpose operation; Step 2.2: Calculate the variance of the residual terms for each subcarrier. Combined into a residual variance matrix ∑ t : Where, diag represents a diagonal matrix with the elements in parentheses as diagonal elements; Step 2.3: Update the channel estimate Collect noisy observations of each subcarrier channel matrix The v-th element is used to form a new row vector of the noisy channel matrix. The model is as follows: in, This represents the v-th row vector of a noise-free sparse channel. For each element in the v-th row and k-th column of a noise-free sparse channel, 1 ≤ k ≤ K; n t,v The noise term follows a standard normal distribution. Step 3: Update residual term V t ; Step 3.1: Calculate the gradient term b t ; Introducing a fully connected layer to update the gradient term b t =[b t [1],…,b t [K] T ; where b of each subcarrier t [k] is determined by the contraction function relative to The gradient is obtained; the intermediate variable of the gradient for each subcarrier k. After the calculation is complete, they are merged into a vector and passed through a fully connected network g. t () to obtain the denoised b t : in, This represents the operation of calculating partial derivatives; Step 3.2: Calculate the residual term V for this layer. t ; With the help of the Onsager correction term b for each subcarrier k t [k]v t-1 [k] to update the residual terms on each subcarrier; after the residual terms of all subcarriers have been calculated, the total residual terms V of all subcarriers in the previous layer are used. t-1 Through the fully connected layer f t To introduce a momentum factor using parentheses (), thus accelerating convergence: in, Let be the residual intermediate variable of the k-th subcarrier. The total residual intermediate variable for all subcarriers; Step 4: After iterating through the TN-layer network, output the complete frequency-sparse domain channel estimation matrix of the TN-th layer. The frequency-antenna domain channel estimation matrix is ​​calculated using the frequency-selective redundancy dictionary D[k], 1≤k≤K.

3. The knowledge-driven deep learning-based fast estimation method for high-dimensional TSN channels as described in claim 2, characterized in that, From China Resumption Design a denoiser based on minimum mean square error. Due to the identical support set property of the channel sparse matrix, for any subcarrier index k, They all have the same sparsity; assuming the noise-free sparse channel vector to be recovered follows a Bernoulli-Gaussian distribution, then The prior distribution is modeled as follows: in, It follows a complex Gaussian distribution, and γ is The probability, ε is hour The variances of these parameters are all trainable parameters. Shrinkage function based on MMSE denoising Represented as: in, Let v be the row vector of the denoised sparse channel estimation matrix. Let φ(x) be the element in the v-th row and k-th column of the denoised sparse channel estimation matrix, 1≤k≤K, φ(x) be the denoising function, and P be a diagonal matrix of trainable prior parameters. when After the calculation is complete, the channel estimate is recombined and updated: in, 1≤v≤V represents the element in the v-th row and k-th column of the denoised sparse channel estimation matrix.

Images