A channel estimation method based on optimized RIS phase shift matrix
Through tensor modeling and parallel factor decomposition model, the RIS phase shift matrix is optimized, and the least squares iterative estimation algorithm is designed, which solves the problem of high overhead and complexity of channel estimation in RIS-assisted wireless communication system, and improves channel estimation performance and system efficiency.
Patent Information
- Application Number
- CN202211361112.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-11-02
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2042-11-02
AI Technical Summary
In RIS-assisted wireless communication systems, there is high overhead and complexity in the channel estimation process, and the existing methods lack optimization when selecting RIS phase shift matrix, affecting channel estimation performance and system efficiency.
The channel estimation method based on tensor modeling and parallel factor decomposition model is adopted. By optimizing the selection of RIS phase shift matrix, the least squares iterative estimation algorithm is designed to reduce system overhead and improve channel estimation performance.
The reconstruction of the channel matrix is realized, which reduces system overhead and complexity, improves channel estimation performance and convergence speed, and ensures the stability of the algorithm.
Smart Images

Figure CN115801501B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of wireless communication technology, and in particular to a channel estimation method based on optimizing a RIS phase shift matrix. Background Art
[0002] Reconfigurable intelligent surface (RIS) is an emerging technology that is expected to achieve ultra-high spectral efficiency in future wireless communications. Specifically, RIS, as a physical layer technology, can be deployed in indoor and outdoor environments to manipulate the propagation of electromagnetic waves by changing the electromagnetic units on the RIS surface. RIS is a rectangular array composed of a large number of controllable elements, each of which can change the phase, amplitude or polarization state of the incident electromagnetic wave. The metasurface manipulates the incident electromagnetic wave received from the transmitter and then retransmits the electromagnetic wave to the receiver. It is worth mentioning that each RIS element can be controlled independently, so high-gain passive beamforming can be achieved by designing the coefficients of these RIS units. Unlike traditional active phased arrays, RIS cannot perform complex signal processing operations. Therefore, RIS does not introduce additional noise when redirecting the propagation of electromagnetic waves. These unique characteristics make this technology stand out from traditional active phased arrays. Generally speaking, RIS usually acts as an intermediary between the transmitter and the receiver to assist in communication. By deploying RIS between the transmitter and the receiver, the wireless propagation channel can be dynamically controlled, which is essentially unique compared to the design of traditional wireless communications that only focuses on the transceiver.
[0003] In recent years, many scholars have applied RIS to the research of wireless communication systems. On the one hand, RIS can actively enrich channel scattering conditions and enhance the multiplexing gain of wireless communication systems; on the other hand, RIS can realize signal propagation direction control and co-phase superposition in three-dimensional space, increase the strength of received signals, and improve the transmission performance between communication devices. Therefore, RIS has great potential for coverage enhancement and capacity improvement of future wireless networks, providing virtual line-of-sight links, eliminating local coverage holes, serving cell edge users, solving co-frequency interference between cells, and thus realizing an intelligent reconfigurable wireless environment. At the same time, RIS can also be combined with many emerging communication technologies, such as physical layer security, millimeter wave communication, drones, index modulation, non-orthogonal multiple access, synchronous wireless information and power transmission, and cell-free massive MIMO. In actual wireless communication systems, RIS has two significant characteristics. First, RIS is composed of passive reflective elements and cannot actively process complex signals; second, RIS is often equipped with a huge number of reflective units, which makes the channel dimension associated with it huge. Therefore, the research on RIS-assisted wireless communication MIMO channel estimation is challenging, and how to design an efficient channel estimation algorithm has also become the research interest of many scholars. Some scholars have proposed a channel estimation method based on minimum mean square error. First, the RIS unit is set to the "OFF" state to estimate the base station-user direct link channel, and then the indirect link base station-RIS-user is divided into multiple sub-stages, in which each RIS element is set to the "ON" state. However, due to the huge number of RIS units, the number of sub-stages will also increase accordingly, which introduces more channel estimation pilot overhead. In order to reduce the CSI estimation overhead, some scholars have proved that the estimation performance depends on the setting of the RIS phase shift matrix, and designed a channel estimation scheme, in which the RIS element takes the discrete Fourier transform (DFT) matrix during the channel estimation, and uses the minimum mean square error estimation results to guide the optimization design. It is shown from theory and simulation that the estimation variance is one order of magnitude smaller than other RIS mode methods. Some scholars have studied the RIS-assisted orthogonal frequency division multiplexing (OFDM) technology under frequency selective channels. First, a new dynamic passive beamforming scheme is proposed to reduce the fading of artificial channels, thereby improving the passive beamforming performance. Then, in order to solve the local optimal solution of the problem, an efficient alternating optimization algorithm is proposed. Finally, experiments show that in the case of dynamic passive beamforming, this method is more dynamic than the case without RIS or fixed passive beamforming, with fewer users served at the same time at each time point, and a more flexible balance between multi-user diversity and RIS passive beamforming gain, thereby improving the common rate. Some scholars pointed out that under the current RIS-assisted communication beamforming design, it is necessary to estimate the (KMN+KM) channel coefficients, where K, N, and M represent the number of users, the number of RIS reflector elements, and the number of BS antennas, respectively.Since each RIS unit sends the transmission signals of all users to the BS, there will be a lot of redundancy in the channel estimation process. In order to use this redundancy to reduce the channel estimation time, relevant scholars proposed a RIS-assisted uplink multi-user communication channel estimation framework based on three-phase pilot. This framework estimates the User-BS direct channel and User-RIS-BS reflected channel of typical users in the first and second stages respectively, while the User-RIS-BS channel of other users is estimated with low overhead in the third stage by utilizing its strong correlation with the channel of the typical user. In addition, some scholars proposed a channel estimation technology based on deep learning (DL). This scheme designs a dual convolutional neural network (CNN) structure and feeds the received pilot signal back to CNN to estimate the direct channel and cascade channel. The performance of the proposed DL method is evaluated and compared with the most advanced DL-based technology, proving its superior performance. Existing channel estimation research usually takes the phase shift matrix of RIS as a local DFT matrix (i.e., selecting the first part of the rows or columns of the DFT matrix) or a sparse matrix. However, in different system models, which rows or columns of the DFT matrix are selected as the required RIS phase shift matrix will have a certain impact on the channel estimation performance. Similarly, if the RIS phase shift matrix is considered to be a sparse matrix, then how to select the values of the sparse elements and how to determine the sparsity of the sparse matrix will also lead to different estimation results brought by different RIS phase shift matrices. Therefore, in MIMO wireless communication systems, in order to achieve a better auxiliary effect of RIS, it is first necessary to optimize the design of the adjustable phase shift matrix, so as to combine different channel estimation methods to achieve the ideal goal. Summary of the invention
[0004] The present invention proposes a channel estimation method based on optimizing the RIS phase shift matrix. While optimizing the RIS phase shift matrix, the channel matrix is reconstructed, which can not only reduce system overhead and complexity, but also has the advantages of improving channel estimation performance, fast convergence speed and good stability.
[0005] In order to achieve the above object, the present invention adopts the following technical solutions:
[0006] A channel estimation method based on optimizing the RIS phase shift matrix comprises the following steps:
[0007] Step A: tensor modeling of the RIS-assisted single-user MIMO uplink communication system;
[0008] Step B: Decomposing the received signal in tensor form according to the tensor and parallel factor decomposition model to obtain a two-dimensional received signal model;
[0009] Step C: Assuming that the RIS reflection unit works in different states, different forms of matrices are used as RIS phase shift matrices, and the matrices are optimized and selected;
[0010] Step D: Designing a channel iterative estimation algorithm based on the least squares method;
[0011] Step E: Repeat step D until the algorithm converges.
[0012] Preferably, the step A specifically comprises:
[0013] Consider a single-cell single-user MIMO uplink communication system, where there is a User with M antennas and a BS with L transmitting antennas. A RIS plane with N adjustable reflection units is placed between the BS and the User. The channel matrices between RIS-BS and User-RIS are defined as and Then in T consecutive time slots, the received signal of BS is expressed as:
[0014] Y k =HΩ k GX+W k ,
[0015] In the above formula represents the observation signal matrix at the BS end, represents the k-th phase configuration state of RIS, represents the phase shift matrix of RIS, Represents T consecutive orthogonal pilot signals that satisfy XX H =I M , Indicates that the mean is 0 and the variance is σ 2 Additive Gaussian noise. Since X satisfies XX H =I M ,exist Multiply both sides of the equal sign by X H , get the observation matrix after removing the pilot signal for:
[0016]
[0017] in, represents the observation matrix after removing the pilot signal and noise, It represents a new noise signal.
[0018] Preferably, the step B specifically comprises:
[0019] The observation matrix after removing the pilot signal satisfies the basic conditions of the parallel factorization model, then The collection forms of the two-dimensional horizontal, lateral and frontal slices of can be expressed as:
[0020] R h =(G T ⊙Φ)H T +W′ h
[0021] R l =(Φ⊙H)G+W′ l
[0022] R f =(H⊙G T )Φ T +W′ f
[0023] in, and Respectively Three slice matrices of and Represent the noise corresponding to the horizontal, lateral and frontal slices respectively.
[0024] Preferably, the step C specifically comprises:
[0025] Step C1: When the RIS reflection unit is in multi-mode, each reflection unit is in working state, and contains information such as its own phase and amplitude. Here, a matrix composed of a partial DFT matrix and a Zadoff-Chu sequence is used as a phase shift matrix.
[0026] Step C2: When the RIS reflection unit is in single mode, each element contains only two states, "on" and "off", which can also be considered as an "ON / OFF" mode. At this time, the phase shift matrix is considered to be a matrix composed of a "0-1" matrix with different sparsity, an LDPC code check matrix and a Gold sequence.
[0027] Preferably, the step C1 specifically comprises:
[0028] When the RIS reflection unit is multi-modal, the DFT matrix and Zadoff-Chu sequence are considered to construct the phase shift matrix Φ.
[0029] Existing research has proven that the DFT matrix is a good choice for the phase shift matrix, which is helpful for the channel estimation performance analysis under the conventional system model. It is often not a square matrix, because although the more phase configuration states K, the more conducive to channel estimation, the existence of too many phase configuration states in the system usually increases the overall overhead cost and complexity of the system. Therefore, it is always hoped to use less phase configuration state information to efficiently complete channel estimation, that is, K≤N. In this case, Φ is usually no longer a square matrix, but a fat matrix with more columns than rows, or called a partial DFT matrix, that is:
[0030]
[0031] Zadoff-Chu sequence is also a set of sequence forms with special properties, which mainly include the following specific properties: 1) This sequence is similar to g -jθ The expression is generated by Euler's formula. Its amplitude is constant. At the same time, the elements in the sequence are projected into a two-dimensional plane. If the real part represents the horizontal axis coordinate and the imaginary part represents the vertical axis coordinate, then all the numbers of the Zadoff-Chu sequence will be presented in a circular range; 2) The correlation between different Zadoff-Chu sequences is constant, and the mutual correlation is very small; 3) The autocorrelation coefficient of the Zadoff-Chu sequence is 0. In other words, many orthogonal sequences can be directly obtained through the Zadoff-Chu sequence. It is also possible to consider using the Zadoff-Chu sequence to construct the phase shift matrix Φ.
[0032] Preferably, the step C2 specifically includes:
[0033] When the RIS reflection unit is in single mode, the sparse "0-1" matrix, the LDPC code check matrix and the Gold sequence are considered to construct the phase shift matrix Φ.
[0034] A sparse matrix is a matrix in which the number of non-zero elements is much smaller than the number of zero elements, and the position distribution of non-zero elements is random. In order to describe the proportion of the number of elements in a sparse matrix, the sparsity κ, that is, the ratio of the number of non-zero elements to the number of all elements in the matrix, is usually used to describe the specific state of the sparse matrix, that is, κ = N1 / MN, where it is assumed that the matrix with dimension M×N contains N1 non-zero elements. Consider "0-1" matrices of different sparsities as the phase shift matrix of RIS. Since the matrix dimension is fixed, the sparsity is controlled by controlling the number of non-zero elements.
[0035] In order to enable the information to have a certain error detection and correction capability at the receiving end, the information is encoded to form codewords, and the basis of encoding is the generator matrix, and the basis of error detection and correction is the check matrix. The full name of LDPC code is low-density parity-check code, which is a type of linear block code with a sparse check matrix. Specifically, the check matrix of LDPC code has many zero elements, and the number often accounts for more than 50% of the total number of matrix elements, and the rest are elements with a value of 1, which is why it can be called low density. When the LDPC code check matrix is generated, the 0 and 1 elements must meet three constraints: 1) The row weight of each row of the check matrix is the same, that is, the number of 1 elements in each row is the same; 2) The column weight of each column of the check matrix is the same, that is, the number of 1 elements in each column is the same; 3) The number of 1 elements in the same position in every two columns of the check matrix does not exceed 1.
[0036] The Gold sequence is obtained by adding two M sequences modulo 2. The M sequence, whose full name is long linear feedback shift register sequence, is a pseudo-random sequence widely used in spread spectrum communication, code division multiple access, and encryption synchronization. In simple terms, the M sequence can be obtained through the primitive polynomial. For example, consider the primitive polynomial as:
[0037] f = a 7 +a 3 +a 2 +3
[0038] At this time, the shift register has 8 levels, and the corresponding length can be generated 8 -1=127 M sequence, and the coefficient of the register is [1 0 1 1 0 0 0 1]. Multiply the values of the register and the coefficient from the 2nd to the 8th bit and sum them, then use the modulo 2 operation, and finally store the result in the lowest bit of the register. As the register continues to shift to the right, the corresponding M sequence can be obtained. The Gold sequence is obtained by adding two M sequences modulo 2. Its main features include: 1) Good randomness and easy to generate. Two 8-level shift registers can generate up to 2 8 +1 = 129 of length 2 8 -1=127 Gold sequence; 2) In terms of correlation, the Gold sequence has better cross-correlation characteristics than the M sequence.
[0039] Preferably, the step D specifically comprises:
[0040] Combined with the least squares method, by initializing one of the unknown matrices, the estimated value of the other matrix is calculated accordingly, and then the estimated value obtained this time is used to estimate the previous unknown matrix. Set reasonable initialization parameter values and bring them into the alternating iterative algorithm framework to estimate the relevant parameters of the sparse channel to achieve reconstruction of the channel matrix.
[0041] Compared with the prior art, the present invention has the following beneficial effects:
[0042] The present invention performs tensor modeling on the RIS-assisted single-user MIMO uplink communication system. According to the tensor and parallel factor decomposition model, the received signal in tensor form is decomposed to obtain a two-dimensional received signal model. It is assumed that the RIS reflection unit works in different states, and different forms of matrices are used as RIS phase shift matrices, which are optimized and selected. Finally, a channel iterative estimation algorithm is designed according to the least squares method. Numerical simulation experiments explore the influence of relevant parameters on the estimation performance, and verify that the proposed algorithm realizes the reconstruction of the channel matrix while optimizing the RIS phase shift matrix, which can not only reduce system overhead and complexity, but also has the advantages of improving channel estimation performance, fast convergence speed and good stability. BRIEF DESCRIPTION OF THE DRAWINGS
[0043] Figure 1 is a flow chart of the present invention;
[0044] Figure 2 Schematic diagram of the variation curve of NMSE with SNR under different partial DFT phase shift matrices;
[0045] Figure 3 Schematic diagram of the variation curve of NMSE with SNR under different sparsity κ;
[0046] Figure 4 Schematic diagram of the variation curve of NMSE with SNR corresponding to different selected phase shift matrices;
[0047] Figure 5 The schematic diagram of the curve of NMSE changing with SNR when K is different;
[0048] Figure 6 Schematic diagram of the variation curve of NMSE with SNR when N is different;
[0049] Figure 7 It is a system structure diagram of the present invention. DETAILED DESCRIPTION
[0050] The technical solutions in the embodiments of the present invention will be described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.
[0051] Example
[0052] Reference Figure 1 , a channel estimation method based on optimizing the RIS phase shift matrix, comprising the following steps:
[0053] Step A: tensor modeling of the RIS-assisted single-user MIMO uplink communication system;
[0054] Step B: Decomposing the received signal in tensor form according to the tensor and parallel factor decomposition model to obtain a two-dimensional received signal model;
[0055] Step C: Assuming that the RIS reflection unit works in different states, different forms of matrices are used as RIS phase shift matrices, and the matrices are optimized and selected;
[0056] Step D: Designing a channel iterative estimation algorithm based on the least squares method;
[0057] Step E: Repeat step D until the algorithm converges.
[0058] In this embodiment, step A specifically includes:
[0059] Consider a single-cell single-user MIMO uplink communication system, where there is a User with M antennas and a BS with L transmitting antennas. A RIS plane with N adjustable reflection units is placed between the BS and the User. The channel matrices between RIS-BS and User-RIS are defined as and Then in T consecutive time slots, the received signal of BS is expressed as:
[0060] Y k =HΩ k GX+W k ,
[0061] In the above formula represents the observation signal matrix at the BS end, represents the k-th phase configuration state of RIS, represents the phase shift matrix of RIS, Represents T consecutive orthogonal pilot signals that satisfy XX H =I M , Indicates that the mean is 0 and the variance is σ 2 Additive Gaussian noise. Since X satisfies XX H =I M ,exist Multiply both sides of the equal sign by x H , get the observation matrix after removing the pilot signal for:
[0062]
[0063] in, represents the observation matrix after removing the pilot signal and noise, It represents a new noise signal.
[0064] In this embodiment, step B specifically includes:
[0065] The observation matrix after removing the pilot signal satisfies the basic conditions of the parallel factorization model, then The collection forms of the two-dimensional horizontal, lateral and frontal slices of can be expressed as:
[0066] R h =(G T ⊙Φ)H T +W′ h
[0067] R l =(Φ⊙H)G+W′ l
[0068] R f =(H⊙G T )Φ T +W′ f
[0069] in, and Respectively Three slice matrices of and Represent the noise corresponding to the horizontal, lateral and frontal slices respectively.
[0070] In this embodiment, step C specifically includes:
[0071] Step C1: When the RIS reflection unit is in multi-mode, each reflection unit is in working state, and contains information such as its own phase and amplitude. Here, a matrix composed of a partial DFT matrix and a Zadoff-Chu sequence is used as a phase shift matrix.
[0072] Step C2: When the RIS reflection unit is in single mode, each element contains only two states, "on" and "off", which can also be considered as an "ON / OFF" mode. At this time, the phase shift matrix is considered to be a matrix composed of a "0-1" matrix with different sparsity, an LDPC code check matrix and a Gold sequence.
[0073] In this embodiment, step C1 specifically includes:
[0074] When the RIS reflection unit is multi-modal, the DFT matrix and Zadoff-Chu sequence are considered to construct the phase shift matrix Φ.
[0075] Existing research has proven that the DFT matrix is a good choice for the phase shift matrix, which is helpful for the channel estimation performance analysis under the conventional system model. It is often not a square matrix, because although the more phase configuration states K, the more conducive to channel estimation, the existence of too many phase configuration states in the system usually increases the overall overhead cost and complexity of the system. Therefore, it is always hoped to use less phase configuration state information to efficiently complete channel estimation, that is, K≤N. In this case, Φ is usually no longer a square matrix, but a fat matrix with more columns than rows, or called a partial DFT matrix, that is:
[0076]
[0077] Zadoff-Chu sequence is also a set of sequence forms with special properties, which mainly include the following specific properties: 1) This sequence is similar to e -jθ The expression is generated by Euler's formula. Its amplitude is constant. At the same time, the elements in the sequence are projected into a two-dimensional plane. If the real part represents the horizontal axis coordinate and the imaginary part represents the vertical axis coordinate, then all the numbers of the Zadoff-Chu sequence will be presented in a circular range; 2) The correlation between different Zadoff-Chu sequences is constant, and the mutual correlation is very small; 3) The autocorrelation coefficient of the Zadoff-Chu sequence is 0. In other words, many orthogonal sequences can be directly obtained through the Zadoff-Chu sequence. It is also possible to consider using the Zadoff-Chu sequence to construct the phase shift matrix Φ.
[0078] In this embodiment, step C2 specifically includes:
[0079] When the RIS reflection unit is in single mode, the sparse "0-1" matrix, the LDPC code check matrix and the Gold sequence are considered to construct the phase shift matrix Φ.
[0080] A sparse matrix is a matrix in which the number of non-zero elements is much smaller than the number of zero elements, and the position distribution of non-zero elements is random. In order to describe the proportion of the number of elements in a sparse matrix, the sparsity κ, that is, the ratio of the number of non-zero elements to the number of all elements in the matrix, is usually used to describe the specific state of the sparse matrix, that is, κ = N1 / MN, where it is assumed that the matrix with dimension M×N contains N1 non-zero elements. Consider "0-1" matrices of different sparsities as the phase shift matrix of RIS. Since the matrix dimension is fixed, the sparsity is controlled by controlling the number of non-zero elements.
[0081] In order to enable the information to have a certain error detection and correction capability at the receiving end, the information is encoded to form codewords, and the basis of encoding is the generator matrix, and the basis of error detection and correction is the check matrix. The full name of LDPC code is low-density parity-check code, which is a type of linear block code with a sparse check matrix. Specifically, the check matrix of LDPC code has many zero elements, and the number often accounts for more than 50% of the total number of matrix elements, and the rest are elements with a value of 1, which is why it can be called low density. When the LDPC code check matrix is generated, the 0 and 1 elements must meet three constraints: 1) The row weight of each row of the check matrix is the same, that is, the number of 1 elements in each row is the same; 2) The column weight of each column of the check matrix is the same, that is, the number of 1 elements in each column is the same; 3) The number of 1 elements in the same position in every two columns of the check matrix does not exceed 1.
[0082] The Gold sequence is obtained by adding two M sequences modulo 2. The M sequence, whose full name is long linear feedback shift register sequence, is a pseudo-random sequence widely used in spread spectrum communication, code division multiple access, and encryption synchronization. In simple terms, the M sequence can be obtained through the primitive polynomial. For example, consider the primitive polynomial as:
[0083] f = a 7 +a 3 +a 2 +3
[0084] At this time, the shift register has 8 levels, and the corresponding length can be generated 8 -1=127 M sequence, and the coefficient of the register is [1 0 1 1 0 0 0 1]. Multiply the values of the register and the coefficient from the 2nd to the 8th bit and sum them, then use the modulo 2 operation, and finally store the result in the lowest bit of the register. As the register continues to shift to the right, the corresponding M sequence can be obtained. The Gold sequence is obtained by adding two M sequences modulo 2. Its main features include: 1) Good randomness and easy to generate. Two 8-level shift registers can generate up to 2 8 +1 = 129 of length 2 8 -1=127 Gold sequence; 2) In terms of correlation, the Gold sequence has better cross-correlation characteristics than the M sequence.
[0085] In this embodiment, step D specifically includes:
[0086] Combined with the least squares method, by initializing one of the unknown matrices, the estimated value of the other matrix is calculated accordingly, and then the estimated value obtained this time is used to estimate the previous unknown matrix. Set reasonable initialization parameter values and bring them into the alternating iterative algorithm framework to estimate the relevant parameters of the sparse channel to achieve reconstruction of the channel matrix.
[0087] In this embodiment, a comparison between the transmission scheme of this embodiment and other existing transmission schemes will be given below to make the advantages and features of this embodiment more obvious.
[0088] Setting of simulation parameters: The iterative algorithm involved is designed based on parallel factorization and least square method, and is named TF-LS-CE (Tensor factorization based least square channel estimation) iterative algorithm. The channel matrices H and G both obey complex Gaussian distribution with mean 0 and variance 1. The maximum number of iterations of TF-LS-CE is set to 20 times, and the threshold ξ=10 -3 All simulations are averages obtained after 1000 Monte Carlo simulations.
[0089] Reference Figure 2 , considering the change of channel estimation performance, namely NMSE with signal to noise ratio (SNR) when the phase shift matrix takes different partial DFT matrices. The simulation parameters involved in the figure are set to K=16, M=N=L=T=32. Here, the phase shift matrix takes three ways to take part of the DFT matrix, namely, taking the first K rows, the last K rows, and randomly taking K rows of the DFT matrix with dimension N×N. By comparing the simulation diagrams, it can be found that taking the first K rows, the last K rows, and randomly taking K rows of the DFT matrix can all achieve good estimation performance, but the first two methods, namely taking the first K rows and the last K rows, have the same NMSE, and the estimation performance corresponding to randomly taking K rows is slightly improved. This shows that when selecting a partial DFT matrix as the phase shift matrix, the channel estimation performance will be affected by different choices. Randomly selecting some rows in the DFT matrix will be beneficial to channel estimation. This is because the columns of the phase shift matrix formed at this time are more uncorrelated, which is more conducive to solving linear channel estimation.
[0090] Reference Figure 3, shows the process of the change of NMSE of channel estimation with the increase of SNR for "0-1" matrices with different sparsities. At this time, K=16, M=N=L=T=32. It can be seen from the figure that the "0-1" phase shift with different sparsities has a great influence on the performance of channel estimation, especially when κ=0.75, the curve value of NMSE is very large, which also shows that the channel estimation performance is poor at this time. As the values of sparsity are getting smaller and smaller when 0.625 and 0.5, the estimation performance of the algorithm is getting better and better. When the sparsity is smaller, the NMSE curve is closer. The figure shows that the NMSE obtained when κ=0.375 is smaller than that when κ=0.25, which means that the estimation performance is slightly better. Since only a part of the sparsity values are selected for simulation experiments, a screening method and idea is only provided when the phase shift matrix uses a "0-1" sparse matrix. Theoretically, the NMSE curve of the estimation algorithm using a sparse phase shift matrix cannot continue to decrease as the sparsity decreases. There will always be a critical value, but when the sparsity κ≤0.3, the estimation algorithm will occasionally not work during the Monte Carlo simulation. Therefore, this experiment considers κ=0.375 as a relatively optimal sparsity, that is, the corresponding "0-1" sparse matrix selected at this time is more conducive to improving the performance of channel estimation.
[0091] Reference Figure 4 , the process of NMSE curve changing with SNR when the phase shift matrix selects partial DFT matrix (random selection of rows), "0-1" sparse matrix (κ=0.5), LDPC check matrix, Zadoff-Chu sequence and Gold sequence is given, and K=16, M=N=L=T=32 are set at the same time. As can be seen from the figure, with the increase of SNR, all curves show a downward state, but it is obvious that the estimation performance corresponding to LDPC check matrix and Gold sequence is poor, especially at low signal-to-noise ratio, the NMSE corresponding to Gold sequence is even greater than 0. In comparison, the NMSE corresponding to Zadoff-Chu sequence is slightly lower than that of sparse matrix, but still when the phase shift matrix selects to use partial DFT matrix, TF-LS-CE iterative algorithm can obtain the best estimation performance. Therefore, partial DFT matrix (random selection of rows) is selected as the result of optimal screening of phase shift matrix.
[0092] Reference Figure 5, shows the law of the NMSE curves of the TF-LS-CE and VAMP-CE algorithms changing with SNR when the number of phase configuration states K of the RIS unit is 8, 16 and 32 respectively, and M=N=L=T=32. According to the simulation experiments in the above text, the NMSE curves of the channel matrices H and G are basically consistent, so only the NMSE curve corresponding to the channel matrix H is shown in the following simulation. As can be seen from the figure, it is obvious that the increasing number of K helps to improve the channel estimation performance. When K=16 and 32, the estimation performance of the two is relatively close, and the TF-LS-CE algorithm has a slight performance advantage in comparison. The key point is that when K=8, the NMSE curve of the TF-LS-CE algorithm is much smaller than that of the VAMP-CE algorithm, and as the signal-to-noise ratio SNR increases, the gap between them is still increasing, which shows that when the number of phase configurations of the RIS unit is small, the TF-LS-CE algorithm has a significant performance improvement over the VAMP-CE algorithm, or in other words, in order to achieve the same channel estimation performance and obtain the same NMSE, the former requires fewer phase configurations of the RIS unit. In the process of channel estimation, although the increase of K is beneficial to the completion of the channel estimation problem, blindly increasing the number of phase configurations of the RIS unit will directly affect the system overhead and complexity. Therefore, while ensuring the channel estimation performance, using fewer phase state configurations will help reduce the overhead and complexity of the entire communication system. Therefore, this conclusion is particularly important for improving system performance, reducing complexity and system overhead.
[0093] Reference Figure 6 , is the performance comparison of the algorithm under different SNRs when different numbers of RIS units are taken, and at this time M=K=L=T=32. As can be seen from the figure, when N is 32 and 48 respectively, the performance of the TF-LS-CE algorithm is significantly better than that of the VAMP-CE algorithm, and when N=32, the NMSE curves corresponding to the two algorithms are significantly lower than the corresponding curves when N=48. This is because as N increases, the number of reflection units of RIS will increase, which will lead to an increase in the dimension of the matrix associated with RIS, and the channel dimension required to be estimated in the channel estimation process will also increase significantly. The complexity of the algorithm will naturally increase, which is not conducive to the channel estimation process, and therefore will cause the channel estimation algorithm to deteriorate as the number of reflection units of RIS increases.
[0094] Reference Figure 7 , is a system structure diagram of this embodiment, which is a system structure diagram of channel estimation based on optimizing RIS phase shift matrix, including: initialization module 1, channel estimation module 2 and channel estimation module 3. Specifically:
[0095] Initialization module 1, initialization setting of relevant parameters and starting the iteration process;
[0096] Channel estimation module 2, based on the initial value, updates the unknown matrix 1;
[0097] Channel estimation module 3 updates unknown matrix 2 based on unknown matrix 1 obtained in the previous module, and substitutes the updated value into the next iteration process.
[0098] The specific initialization and channel estimation process in the channel estimation method based on optimizing the RIS phase shift matrix can be referred to the above embodiment, and the embodiments of the present invention will not be described in detail here.
[0099] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.
Claims
1. A channel estimation method based on optimizing the RIS phase shift matrix, characterized in that: The following steps are involved: Step A: tensor modeling of the RIS-assisted single-user MIMO uplink communication system; Step B: Decomposing the received signal in tensor form according to the tensor and parallel factor decomposition model to obtain a two-dimensional received signal model; Step C: The RIS reflector unit works in different states, uses matrices of different forms as RIS phase shift matrices, and optimizes and selects them; specifically: the RIS reflector unit uses different phase shift matrices in two working states, namely, multi-mode and single-mode working states, and optimizes and selects the phase shift matrix, which specifically includes the following steps: Step C1: When the RIS reflection unit is in multi-mode, each reflection unit is in working state and contains respective phase and amplitude information. Here, a matrix composed of a partial DFT matrix and a Zadoff-Chu sequence is used as a phase shift matrix. Step C2: When the RIS reflection unit is in single mode, each element contains only two states, "on" and "off", that is, "ON / OFF" mode. At this time, the phase shift matrix is a matrix composed of "0-1" matrices with different sparsities, LDPC code check matrices and Gold sequences; Step D: Designing a channel iterative estimation algorithm based on the least squares method; Step E: Repeat step D until the algorithm converges.
2. A channel estimation method based on optimizing the RIS phase shift matrix according to claim 1, characterized in that: In step A, a single-cell single-user MIMO uplink communication system has a User with M antennas and a BS with L transmitting antennas. A RIS plane with N adjustable reflection units is placed between the BS and the User. The channel matrices between RIS-BS and User-RIS are defined as: and Then in T consecutive time slots, the received signal of BS is expressed as: Y k =HΩ k GX+W k , In the above formula represents the observation signal matrix at the BS end, represents the kth phase configuration state of RIS, represents the phase shift matrix of RIS, Represents T consecutive orthogonal pilot signals that satisfy XX H =I M , Indicates that the mean is 0 and the variance is σ 2 Additive Gaussian noise; since X satisfies XX H =I M ,exist Multiply both sides of the equal sign by X H , get the observation matrix after removing the pilot signal for: in, represents the observation matrix after removing the pilot signal and noise, It represents a new noise signal.
3. A channel estimation method based on optimizing the RIS phase shift matrix according to claim 1, characterized in that: In step B, the observation matrix after removing the pilot signal satisfies the basic conditions of the parallel factor decomposition model. The collection forms of the two-dimensional horizontal, lateral and frontal slices of can be expressed as: R h =(G T ⊙Φ)H T +W' h R l =(Φ⊙H)G+W' l R f =(H⊙G T )F T +W' f in, and Respectively Three slice matrices of and Represent the corresponding noise in horizontal, lateral and frontal slices respectively.
4. A channel estimation method based on optimizing the RIS phase shift matrix according to claim 1, characterized in that: The step C1 specifically comprises the following steps: When the RIS reflection unit is multi-modal, the DFT matrix and Zadoff-Chu sequence are used to construct the phase shift matrix Φ; Phase Shift Matrix It is not a square matrix, and uses less phase configuration state information to efficiently complete channel estimation, that is, K≤N. In this case, Φ is no longer a square matrix, but a fat matrix with more columns than rows, or called a partial DFT matrix, that is:
5. The channel estimation method based on optimizing the RIS phase shift matrix according to claim 1, characterized in that: The step C2 specifically includes the following steps: When the RIS reflector unit is in single mode, it uses a sparse "0-1" matrix, an LDPC code check matrix and a Gold sequence to construct a phase shift matrix Φ; A sparse matrix means that among all the elements of the matrix, the number of non-zero elements is much smaller than the number of zero elements, and the position distribution of non-zero elements is random. In order to describe the proportion of the number of elements in a sparse matrix, the sparsity κ, that is, the ratio of the number of non-zero elements to the number of all elements in the matrix, is used to describe the specific state of the sparse matrix, that is, κ = N1 / MN, where a matrix with a dimension of M×N contains N1 non-zero elements. "0-1" matrices with different sparsities are used as the phase shift matrix of RIS. Since the matrix dimension is fixed, the sparsity is controlled by controlling the number of non-zero elements. The information is encoded to form codewords, and the basis of encoding is the generator matrix, and the basis of error detection and correction is the check matrix; the full name of LDPC code is low-density parity-check code, which is a type of linear block code with a sparse check matrix; The Gold sequence is obtained by adding two M sequences modulo 2. The M sequence, whose full name is long linear feedback shift register sequence, is a pseudo-random sequence widely used in spread spectrum communication, code division multiple access and encryption synchronization. The M sequence can be obtained through the primitive polynomial.
6. A channel estimation method based on optimizing the RIS phase shift matrix according to claim 1, characterized in that: In the step D, the least squares method is combined to initialize one of the unknown matrices, calculate the estimated value of the other matrix accordingly, and then use the estimated value obtained this time to estimate the previous unknown matrix; set reasonable initialization parameter values, and bring them into the alternating iterative algorithm framework to estimate the relevant parameters of the sparse channel to achieve reconstruction of the channel matrix.
