Hybrid precoding method based on IRS-aided MIMO system

Abstract

The application provides an IRS-assisted MIMO system hybrid precoding method, and mainly solves the problems of high power consumption and high calculation complexity of the prior art hybrid precoding, and the implementation scheme is as follows: a hybrid precoding model of the IRS-assisted MIMO system is established; a whole spectral efficiency R formula of the hybrid precoding model is constructed; a deep neural network DNN is established, a reflection matrix dataset under different channel environments is trained offline, and the reflection matrix with the optimal spectral efficiency R is selected from the dataset in the online deployment stage; according to the optimized reflection matrix, an equivalent channel matrix is obtained, and the optimal hybrid precoding under the channel condition is calculated. The application constructs an analog precoding module based on the PMA structure, and the performance of the analog precoding module is better and the power consumption is lower than those of a traditional phase shifter, and the reflection matrix of the IRS is solved based on the deep learning network DNN, so that the calculation complexity is reduced, and the application can be used in a low-latency real-time communication system.

Description

Technical Field

[0001] This invention belongs to the field of wireless communication technology, and specifically relates to a hybrid precoding method that can be used in low-latency real-time MIMO communication systems. Background Technology

[0002] With the exponential increase in data traffic demand in the 5G era, millimeter-wave communication, offering high speeds, is gaining popularity. While millimeter-wave frequencies offer higher transmission rates, their poor diffraction and penetration capabilities mean that signal interruptions can occur when obstacles block transmission. This critical issue severely impacts the reliability of wireless transmission. Ultra-dense networks (UDNs) offer a solution to improve reliability, but their linear scaling of circuitry, coupled with expensive RF links and high network power consumption, makes the hardware costs of large-scale MIMO systems prohibitively high. Furthermore, the addition of numerous active components to the wireless network, such as cell base stations, relays, and remote radio heads, can lead to severe interference. Therefore, finding low-cost and energy-efficient technologies is crucial for the sustainable development of future communication networks.

[0003] Intelligent reflective resonant relays (IRS) have garnered significant attention in academia as a potential low-cost, energy-efficient communication solution. An IRS consists of numerous reconfigurable passive reflective elements. Each element, under the control of the IRS controller, adjusts the amplitude and phase of the incident signal before reflecting it, thereby reconfiguring the wireless transmission environment. This allows signals to bypass obstacles and establish virtual links between base stations and users. Compared to traditional active repeaters, this intelligent reflective IRS better meets the requirements for low energy consumption. Therefore, research on hybrid precoding techniques in IRS-assisted communication systems holds great promise.

[0004] Nanjing Skoma Electronics Technology Co., Ltd. proposed a "Method for Optimizing the Achievable Combined Rate of a Smart Reflector MIMO System" in its patent application number 202210137646.9. The implementation scheme includes: (1) inputting the number of base station antennas, the number of users, the number of user antennas, the number of smart reflector reflector units, and the channel matrix; (2) using a fractional programming algorithm, the original problem of maximizing the reachable combined rate is transformed into a solvable problem by introducing auxiliary variables; (3) using an alternating maximization algorithm, the diagonal reflection matrix, the combiner, and the precoder that maximize the equivalent objective equation value are calculated; (4) the corresponding maximum system reachable combined rate value is output. Although this method solves the problem of easy attenuation in traditional millimeter-wave MIMO systems through IRS and improves the system reachable combined rate, this method still has two shortcomings: First, the base station analog precoder part is implemented by a phase shifter, which has high power consumption and limited quantization accuracy, resulting in quantization loss; second, in the process of solving the hybrid precoder and reflection matrix, a large number of iterative processes are required, resulting in high coding complexity, which does not meet the requirements of low-latency communication. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of the prior art by proposing a hybrid precoding method for MIMO systems based on IRS assistance, which reduces the power consumption of the analog precoding part of the base station and reduces the coding complexity by using a deep neural network (DNN) to meet the requirements of low-latency communication.

[0006] To achieve the above objectives, the technical solution of the present invention includes the following steps:

[0007] (1) Constructing an IRS-assisted MIMO hybrid precoding model:

[0008] (1a) Establish an analog precoding module F composed of phase modulation array (PMA) structure respectively. RF Analog combiner W RF ;

[0009] (1b) Select the existing digital precoding module F BB and the radio frequency link, connecting it to the analog precoding module F RF Cascaded and combined with the transmitting antenna to form a base station;

[0010] (1c) Establish a system including a receiving antenna and an analog combiner W RF RF link, digital combiner W BB The user terminal consists of modules;

[0011] (1d) The signal at the base station is reflected to the user terminal via the intelligent reflective surface IRS;

[0012] (2) Construct the formula for calculating the overall spectral efficiency R of the MIMO hybrid precoding model:

[0013] (2a) N at the base station s The data stream S is transmitted to the digital precoding module F. BB Processed, the signal reaches the analog precoding module F via the radio frequency link. RF After analog precoding processing, the transmitted signal x = F is generated. RF F BB S is transmitted to the intelligent reflector IRS via the antenna;

[0014] (2b) The IRS reflects the signal generated by the base station to the user by adjusting the reflection coefficient matrix Θ;

[0015] (2c) The user's antenna receives the reflected signal y from the IRS:

[0016]

[0017] Where G represents the channel matrix between the base station and the IRS, H r This represents the channel matrix between the IRS and the user, where ρ represents the received power and n represents the noise signal.

[0018] (2d) Send the reflected signal y to the analog combiner W RF Analog decoding and precoding are performed, and the signal reaches the digital combiner W via the radio frequency link. BB Perform digital decoding and precoding to recover the data stream.

[0019]

[0020] in, It is W RF The conjugate transpose of the matrix. It is W RF The conjugate transpose of ;

[0021] (2e) Based on Shannon's formula and the recovered data stream The formula for calculating the spectral efficiency R of the MIMO hybrid precoding model is as follows:

[0022]

[0023] Where, σ 2 The variance of the noise, and the equivalent channel H eff =H r ΘG, Hybrid precoding matrix F = F RF F BB W = W RF W BB R n =W H W;

[0024] (3) Construct a deep neural network (DNN) and solve for the IRS reflection matrix Θ in the spectral efficiency R:

[0025] (3a) Construct a deep neural network (DNN) consisting of an input layer, a Flatten layer, five fully connected layers, a Lambda layer, and an output layer connected in sequence.

[0026] (3b) Channel combination information H c =H r G is used as the input to the DNN network;

[0027] (3c) Define the loss function of a deep neural network (DNN) as follows:

[0028]

[0029] Where ρ represents the received power, N s This represents the number of data streams transmitted simultaneously, where N is the total number of training samples, and Σ is the equivalent channel H. eff The diagonal matrix after singular value decomposition, taking the first N of Σ s Rows, first N s The columns form a submatrix Σ1;

[0030] (3d) In the offline training phase, the generated dataset {H} c The input is fed into the DNN network, and the weights and biases of the neurons are continuously adjusted and optimized through backpropagation until the loss function is minimized, thus obtaining the reflection matrix dataset {Θ} after offline training.

[0031] (3e) During the online deployment phase, the channel information H' c By inputting the data into a DNN network, the DNN-optimized reflection matrix Θ′ can be obtained.

[0032] (4) Design a hybrid precoding:

[0033] (4a) Calculate the equivalent channel H′ based on the reflection matrix Θ′ optimized by DNN. eff And perform singular value decomposition on it, i.e., H′ eff =U′Σ′V′ H Where U is the combiner matrix and V′ is the precoding matrix;

[0034] (4b) Take the first N elements of matrix V′ s The columns form a submatrix V′1, which is used as the optimal all-digital precoder, i.e., F. opt =V′1, minimize F opt With F RF F BB The Euclidean distance between them is used to solve for the hybrid precoder F. RF F BB :

[0035]

[0036]

[0037]

[0038] in, F is something that PMA can achieve. RF The set of feasible solutions.

[0039] Compared with the prior art, the present invention has the following advantages:

[0040] 1. This invention addresses IRS-assisted communication systems by employing a PMA-based hybrid precoding structure at the base station. Compared to existing phase shifters, this structure consumes less energy and can achieve any phase between [0, 2π] without phase quantization. Furthermore, the PMA structure achieves the same spectral performance as the all-digital structure, which is better than that of the phase shifter. Additionally, the construction of an IRS-assisted PMA-based hybrid precoding system model further improves system energy efficiency.

[0041] 2. Because this invention designs a reflection matrix of the IRS based on a deep learning network (DNN), only the offline-trained reflection matrix needs to be loaded during the online deployment phase. The base station and the IRS do not need to exchange information frequently, which reduces the computational complexity and makes it more suitable for low-latency real-time communication systems. Attached Figure Description

[0042] Figure 1 This is a flowchart illustrating the implementation of the present invention;

[0043] Figure 2 This is a schematic diagram of the IRS-assisted MIMO hybrid precoding model structure constructed in this invention;

[0044] Figure 3 This is a diagram of the DNN structure model for solving the reflection matrix in this invention;

[0045] Figure 4 For the precoding F of the PMA structure simulation in this invention RF A schematic diagram of the feasible region.

[0046] Specific implementation steps

[0047] The present invention will now be described in further detail with reference to the accompanying drawings and specific examples.

[0048] Reference Figure 1 The implementation steps for this example are as follows:

[0049] Step 1: Establish a hybrid precoding system model for the IRS-assisted MIMO system.

[0050] This step is an improvement on the hybrid precoding of existing MIMO systems, namely, using a phase modulation array (PMA) instead of a phase shifter to implement analog precoding and analog combiner, in order to reduce the power consumption of analog precoding.

[0051] Reference Figure 2 The system is constructed as follows:

[0052] 1.1) Establish analog precoding module F RF and analog combiner W RF :

[0053] Analog precoding module F RF and analog combiner W RF All are built on the existing phase modulation array (PMA) structure, which consists of multiple antennas and multiple sets of delay lines. Each antenna is connected to a set of delay lines, and the phase shift values ​​of the delay lines are 0, π / 2, π, and 3π / 2, respectively. Within one modulation cycle, the FPGA controls the RF switch to sequentially turn on two adjacent delay lines. By adjusting the turn-on time, an equivalent amplitude phase excitation can be obtained at the center frequency. Compared with the phase shifter, this structure has no phase resolution limitation and increases the amplitude degree of freedom, and is not subject to constant mode constraints.

[0054] This step consists of the analog precoding module F, which is composed of a phase modulation array (PMA) structure. RF The antennas consist of 64 wires and the delay line consists of 4 groups; the analog combiner W RF The analog precoding module F has 16 antennas and 4 delay line groups. RF and analog combiner W RF All are configured as fully connected structures, connecting each antenna to other sets of delay lines to enable multi-stream data transmission;

[0055] 1.2) Select the existing digital precoding module F BB and the radio frequency link, connecting it to the analog precoding module F RF Cascaded and combined with the transmitting antenna to form a base station;

[0056] 1.3) Establish a system including a receiving antenna and an analog combiner W. RF RF link, digital combiner W BB The user terminal consists of modules;

[0057] 1.4) The signal reaches the receiver through an intelligent reflective surface (IRS) equipped with M reflective elements. The IRS controller reflects the incident signal back to the user by adjusting the reflection coefficient matrix Θ. In this example, the number of reflective elements is set to M = 256.

[0058] Step 2: Construct the formula for the overall spectral efficiency R of the MIMO hybrid precoding model.

[0059] 2.1) N at the base station end s The data stream S is transmitted to the digital precoding module F. BB Processed, the signal reaches the analog precoding module F via the radio frequency link. RF After analog precoding processing, the transmitted signal x = F is generated. RF F BB S is then transmitted via an antenna to the intelligent reflector IRS. In this example, N is taken as, but is not limited to, N. s =4;

[0060] 2.2) The IRS reflects the signal generated by the base station to the user by adjusting the reflection coefficient matrix Θ;

[0061] 2.3) The user's antenna receives the reflected signal y from the IRS:

[0062]

[0063] Where G represents the channel matrix between the base station and the IRS, H r This represents the channel matrix between the IRS and the user, where ρ represents the received power and n represents the noise signal.

[0064] 2.4) Send the reflected signal y to the analog combiner W. RF Analog decoding and precoding are performed, and the signal reaches the digital combiner W via the radio frequency link. BB Perform digital decoding and precoding to recover the data stream.

[0065]

[0066] in, It is W RF The conjugate transpose of the matrix. It is W RF The conjugate transpose of ;

[0067] 2.5) Based on Shannon's formula and the recovered data stream The formula for calculating the spectral efficiency R of the MIMO hybrid precoding model is as follows:

[0068]

[0069] Where, σ 2 H represents the variance of the noise. eff =H r ΘG is the equivalent channel matrix, F = F RF F BB For the hybrid precoding matrix, W = W RF WBB Let R be the hybrid combiner matrix. n =W H W is an intermediate variable, and the reflection matrix Θ and the hybrid precoding matrix F are the variables that need to be solved.

[0070] Step 3: Maximize the spectral efficiency R.

[0071] To facilitate solving the reflection matrix and hybrid precoding, the optimization objective needs to be set as maximizing the system's spectral efficiency. Therefore, the system's spectral efficiency R is modeled as:

[0072]

[0073]

[0074]

[0075]

[0076] in, F is something that PMA can achieve. RF feasible domain, such as Figure 4 As shown in the square area, Due to the constraint of total transmit power, N s This represents the number of data streams transmitted simultaneously. Due to joint optimization F... BB F RF W BB W RF Furthermore, it has non-convex constraints, making it difficult to find the global optimal solution.

[0077] To reduce the difficulty of solving the problem, the objective problem is decomposed into two independent sub-problems: first, the design of the IRS reflection matrix Θ is considered to obtain the equivalent channel H. eff H r ΘG, then utilize H eff Next, we solve for the hybrid precoding.

[0078] Step 4: Construct a deep neural network to solve for the reflection matrix Θ that maximizes the spectral efficiency R.

[0079] 4.1) Setting up the DNN network structure

[0080] Reference Figure 3 This DNN structure model has 9 layers, of which:

[0081] The first layer is the input layer, used to input combined information of the base station-IRS and IRS-user channels.

[0082] The second layer is the Flatten layer, which flattens the three-dimensional channel information into one-dimensional data for transmission to the fully connected layer.

[0083] Layers three through seven consist of five fully connected layers used to compute the reflection matrix Θ to be optimized. The number of neurons from front to back are 4096, 2048, 1024, 512, and 256, respectively. To prevent gradient vanishing or exploding, a batch normalized (BN) structure is set in front of each fully connected layer, and the activation function of the first four fully connected layers is set to the ReLU function. The last fully connected layer does not use any activation function.

[0084] The eighth layer is a custom Lambda layer used to output complex values ​​with constant modulus constraints. Because the IRS involved in this example only adjusts the phase of the incident signal, and the amplitude reflection coefficients are all 1, the reflection matrix Θ has constant modulus constraints. Furthermore, since the elements in the reflection matrix Θ are complex numbers, and the TensorFlow network framework used in this example does not support the processing of complex numbers, a Lambda layer is needed to handle these complex numbers.

[0085] Let vector θ in If the input to the Lambda layer is a real value, then the constant modulus complex value of the output of the Lambda layer is:

[0086] θ out =cos(θ) in )+jsin(θ in );

[0087] The ninth layer is the output layer, used to output the obtained reflection matrix Θ.

[0088] 4.2) Generate a dataset {H} containing channel combination information. c}:

[0089] Channel combination information H c =H r G, as the input to the DNN network, can establish a line-of-sight (LoS) link between the base station and the IRS, and between the IRS and the user, thanks to the use of the intelligent reflector IRS. Therefore, G and H... r Both can employ millimeter-wave channel models with one line-of-sight (LoS) link and multiple non-line-of-sight (NLoS) links, and their generation formulas are expressed as follows:

[0090]

[0091]

[0092] Where α1 and β1 are the complex gains of the line-of-sight paths at the transmitter and receiver, respectively. The complex gains of the non-line-of-sight paths at the transmitter and receiver are respectively, a t a r These are the normalized transmit and receive array response vectors, respectively. These are the azimuth angles for transmitting and receiving signals, respectively. L1 and L2 are the elevation angles of the transmitted and received signals, respectively, and L1 and L2 are the number of propagation paths for the transmitted and received signals, respectively. N t N r These represent the number of transmitting antennas and the number of receiving antennas, respectively.

[0093] In this example, both the transmitter and receiver employ a millimeter-wave channel model with one LoS path and L-1 NLoS paths, where L = L1 = L2 = 30. Channel matrices G and H are generated according to the formulas described above. r The sample, thus obtaining H c To enable the DNN network to better adapt to complex channel environments, 640,000 sets of H were generated. c As a dataset {H c}

[0094] 4.3) Based on the spectral efficiency R, define the loss function for the deep neural network (DNN):

[0095] Since it is difficult to find the optimal reflection matrix Θ as the label, the inverse of the spectral efficiency is used as the loss function. To simplify the expression of the loss function, the optimal all-digital precoding F is first used. opt Alternative Hybrid Precoding F BB F RF Then use the optimal combiner W opt Alternative Hybrid Combiner W BB W RF At this point, the spectral efficiency can be expressed as:

[0096] in

[0097] For the equivalent channel H eff Perform singular value decomposition, i.e., H eff UΣV H Where Σ is the equivalent channel H eff The diagonal matrix after singular value decomposition, U is the combiner matrix, and V is the precoding matrix;

[0098] Take the first N of matrices V and U respectively. s The columns form submatrices V1 and U1, which serve as the optimal all-digital precoder F. opt With the optimal all-digital combiner W opt That is, F opt =V1,W opt U1,

[0099] Due to the sparsity of millimeter-wave channels, H effIt usually has a low rank, so it can be solved by keeping only the strongest N. s To approximate H using a component eff H eff ≈U1Σ1V1, where Σ1 is formed by the first N elements of the matrix Σ. s row, first N s The submatrix composed of columns ultimately yields the expression for the spectral efficiency R', which is only related to the reflection matrix Θ:

[0100]

[0101] Based on the above expression for the spectral efficiency R', the loss function of the DNN is defined as follows:

[0102]

[0103] Where N is the total number of training samples, N s For the number of data streams, this example sets N = 640000. s =4.

[0104] 4.4) Offline Training of Deep Neural Networks (DNNs)

[0105] During the offline training phase, the generated dataset {H} c The input is fed into the DNN network, and the weights and biases of the neurons are continuously adjusted and optimized through backpropagation according to the loss function loss, so as to minimize the loss function and obtain the reflection matrix dataset {Θ} after offline training.

[0106] 4.5) During the online deployment phase, the online channel information H' c The input is fed into the DNN network, which selects the optimal reflection matrix Θ′ based on the reflection matrix dataset {Θ} trained offline, which greatly reduces the computational complexity.

[0107] Step 5, Design Hybrid Precoding

[0108] Based on the DNN-optimized reflection matrix Θ′, the equivalent channel H′ is obtained through singular value decomposition. eff To obtain the optimal all-digital precoding F opt At this point, the design of hybrid precoding can be simplified to minimizing F. opt and F RF F BB The Euclidean distance between them is:

[0109]

[0110]

[0111]

[0112] Finally, the hybrid precoding F is solved. RF F BB .

Claims

1. A hybrid precoding method for MIMO systems based on IRS assistance, characterized in that, Including the following: (1) Constructing an IRS-assisted MIMO hybrid precoding model: (1a) Establish an analog precoding module composed of phase modulation array (PMA) structures. Analog combiner The phase modulation array (PMA) includes radio frequency switches, delay lines, and antennas, wherein each antenna is connected to a set of delay lines, and the phase shift values ​​of the delay lines are 0. , , Furthermore, within one modulation cycle, the FPGA controls the RF switch to sequentially turn on two adjacent delay lines. By adjusting the turn-on time, an equivalent amplitude and phase excitation is obtained at the center frequency. (1b) Select an existing digital precoding module and the radio frequency link, connecting it to the analog precoding module. Cascaded and combined with the transmitting antenna to form a base station; (1c) Establish a system including a receiving antenna and an analog combiner. RF links, digital combiners The user terminal consists of modules; (1d) The signal at the base station is reflected to the user terminal via the intelligent reflector IRS; (2) Construct the formula for the overall spectral efficiency R of the MIMO hybrid precoding model: (2a) Base station end Road data stream Transmitted to digital precoding module Processed, the signal reaches the analog precoding module via the radio frequency link. After analog precoding processing, the transmitted signal is generated. It is transmitted to the intelligent reflector IRS via the antenna; (2b) IRS adjusts the reflection coefficient matrix The base station reflects the signal generated by the base station back to the user; (2c) The user's antenna receives the reflected signal from the IRS. : ； in, This represents the channel matrix between the base station and the IRS. This represents the channel matrix between the IRS and the user. Indicates the received power. Indicates a noise signal; (2d) The reflected signal Send to analog combiner The analog precoding process is performed, and the signal reaches the digital combiner via the radio frequency link. Perform digital decoding and precoding to recover the data stream. : ； in, yes The conjugate transpose of the matrix. yes The conjugate transpose of ; (2e) Based on Shannon's formula and the recovered data stream The spectral efficiency of the MIMO hybrid precoding model is obtained. Calculation formula: ； in, The variance of the noise, equivalent channel Hybrid precoding matrix , , ; (3) Construct a deep neural network (DNN) and solve for the spectral efficiency. The IRS reflection matrix in : (3a) Construct a deep neural network (DNN) consisting of an input layer, a flattened layer, five fully connected layers, a lambda layer, and an output layer connected sequentially; the structural parameters and functions of the five fully connected layers and the lambda layer are as follows: The number of neurons in the five fully connected layers is as follows: , , , and ,in, The number of IRS reflection units is denoted by . The activation function of the first four fully connected layers is the ReLU function, and the last fully connected layer does not use an activation function. In addition, there is a batch normalization (BN) layer in front of each fully connected layer to prevent gradient vanishing or exploding. The Lambda layer is used to process the input real values. Output converted to constant modulus complex value : ； (3b) Combining channel information As input to the DNN network; (3c) Define the loss function of a deep neural network (DNN) as follows: ； in, Indicates the received power. Indicates the number of data streams sent simultaneously. The total number of training samples. For equivalent channel The diagonal matrix after singular value decomposition, take The former row, before Columns form submatrix ; (3d) In the offline training phase, the generated dataset will be... The input is fed into a DNN network, and the weights and biases of the neurons are continuously adjusted and optimized through backpropagation until the loss function is minimized, thus obtaining the reflection matrix dataset after offline training. ; (3e) During the online deployment phase, channel information will be... Inputting the data into a DNN network yields the DNN-optimized reflection matrix. ; (4) Design of hybrid precoding: (4a) Based on the reflection matrix optimized by DNN Calculate the equivalent channel And perform singular value decomposition on it, that is ,in, Let be the combiner matrix. This is the precoding matrix; (4b) Take The front of the matrix Columns form submatrix This is taken as the optimal all-digital precoding, i.e. , minimize and The Euclidean distance between them is used to solve for the hybrid precoder. : ； in, This is something that PMA can achieve. The set of feasible solutions.

2. The method according to claim 1, characterized in that, Channel combination information in step (3a) The channel matrix between the base station and the IRS in Channel matrix between IRS and users The calculation is as follows: ； ； in, , These are the complex gains of the line-of-sight paths at the transmitting and receiving ends, respectively. , These are the complex gains of the non-line-of-sight paths at the transmitting and receiving ends, respectively. , These are the normalized transmit and receive array response vectors, respectively. , These are the number of propagation paths for the transmitted and received signals, respectively. , These represent the number of transmitting antennas and the number of receiving antennas, respectively.

Images