OFDM system channel estimation design method based on LISTA complex network
By adopting the design method of LISTA complex network in OFDM system, the shortcomings of deep learning in complex channel estimation are solved, and channel estimation with higher accuracy and noise anti-noise performance are achieved, and virtual and real separation and iterative threshold algorithms are used to improve wireless communication. channel recovery capability.
Patent Information
- Application Number
- CN202310685915.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-06-12
- Publication Date
- 2024-12-13
AI Technical Summary
Deep learning methods cannot directly deal with complex problems, resulting in insufficient performance of OFDM system channel estimation in wireless communications, especially in terms of spectrum resource occupancy and noise robustness.
The design method based on LISTA complex network is adopted, and the complex signal is processed through virtual and real separation method, represented as logically different real-valued entities, and channel estimation is performed in the network using iterative threshold algorithm and recurrent neural network structure, using small Batch gradient descent algorithm and Adam optimization method train network parameters.
It achieves higher channel estimation accuracy and noise anti-noise performance, and the superiority of the network is verified through simulation experiments, saving network resources and improving channel recovery capabilities.
Smart Images
Figure CN119135480A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of digital signal processing and provides a design method for OFDM system channel estimation based on LISTA complex network. Background Art
[0002] Wireless communication has been booming in recent decades and has been integrated into every aspect of human life. Due to the explosive growth of mobile data traffic, mobile communication systems need to further speed up the transmission speed of data traffic. In order to cope with the rapid growth of mobile communication system's demand for transmission rate, many new technologies have emerged in the development of mobile communication systems.
[0003] In recent years, deep learning has been successfully applied to wireless communication systems, such as channel estimation, channel state information feedback and signal detection, etc. The application of neural networks in OFDM communications is becoming more and more promising. For channel estimation, deep learning methods have not been fully studied. Therefore, in view of the insufficient estimation performance of traditional channel estimation methods in wireless environments, channel estimation methods based on recurrent neural networks are studied. At present, most of the building blocks, techniques and architectures of deep learning are based on real-valued operations and representations. However, recent research on recurrent neural networks and earlier basic theoretical analysis have shown that complex numbers may have richer representation capabilities and can also promote noise-robust memory retrieval mechanisms. The representation function based on complex numbers has begun to receive more and more attention. The implementation of complex networks is easier to optimize, has better generalization characteristics and faster learning capabilities. Therefore, the present invention proposes a complex network based on LISTA for OFDM system channel estimation, which has higher channel estimation accuracy and noise resistance. Summary of the invention
[0004] The purpose of the present invention is to solve the problem that deep learning methods cannot directly process complex numbers, occupy as few spectrum resources as possible, improve the ability of channel estimation, and restore the original data more accurately. A new design method is provided for channel estimation of OFDM systems in wireless communications - a method for channel estimation of OFDM systems based on LISTA complex networks can be designed.
[0005] The specific steps of the design method of OFDM system channel estimation based on LISTA complex network provided by the present invention are as follows:
[0006] First, according to the design requirements, in the OFDM system, the number of subcarriers is N, the channel length is L, and the number of pilots is P, which are located at subcarriers q1, q2, ...q p The transmission symbol of the transmitting antenna on the pilot subcarrier is Then the pilot observation vector at the receiving end can be expressed as:
[0007] y=diag(x)F L h+n (1)
[0008] In the formula, is the received signal of the receiving antenna at the pilot position, is the impulse response vector of the multipath channel from the base station transmitting antenna to the user. The channel length is L, the number of non-zero positions is K, and K<<L is satisfied. That is, the time domain channel presents sparse characteristics, and most of the energy is concentrated on a small number of channel taps. F is the discrete Fourier transform matrix of N points. is a submatrix consisting of the first L columns of the Fourier transform matrix F and the P rows corresponding to the pilot positions, Denotes the additive white Gaussian noise of the channel, let A = diag(x)F L , then the channel estimation problem model (1) can be transformed into the solution of the minimum l0 norm problem, expressed as follows:
[0009]
[0010] Since the receiving vector y, the observation matrix A, the channel vector h and the noise vector n in the OFDM system are all in complex form, most of the current deep learning architectures are based on real-valued operations and representations, and deep learning cannot directly process complex numbers. This design uses the virtual-real separation method under the complex network to process the real part and the imaginary part respectively, and then combine the processed real and imaginary parts to represent the complex number;
[0011] Second, the traditional iterative threshold algorithm process is
[0012]
[0013] where z t is the input vector of the tth iteration, β is the step size of the gradient update, and h t+1 is the output vector of the tth iteration, [η st (z t ;λ)] j =sgn(z t [j])max{|z t [j]|-λ,0}(t=0,1,2,...) represents the element-by-element soft threshold shrinkage function, r j Represents vector z t The jth element in h is updated using the current gradient in each iteration. t Get z t ;
[0014] Third, expand the iterative threshold algorithm into a recurrent neural network and build a complex network LISTA with training parameters. The total number of layers of the network is recorded as T. In the formula, each complex signal has a real component and an imaginary component. The real and imaginary parts of the complex number are separated and represented as logically different real-valued entities. The real-valued algorithm is used internally to simulate complex arithmetic. Let B = βA T , and the following iterative process is obtained:
[0015]
[0016] in and and B r and B i ,y r and i , A r and A i Represents z t 、h t , B, y, the real and imaginary components of A, Represents vector h t The L1 norm of B r , B i , α t (t=0, 1, ... 2T-1) are the parameters that need to be learned in the complex network, is the threshold shrinkage parameter; and is the input vector of the t-th layer network, h t+1 is the output vector of the t-th layer network. The initial value h0 is a zero vector. The h of the last layer output is t This is the channel estimation value h, that is, h = h T ;
[0017] Fourth, use the training data set to train the parameters of the complex network; the main idea is to train a neural network with a specific architecture and fixed depth to produce the best approximation of the sparse code and learn the parameters B t The value of each layer and the shrinkage function threshold λ t All through the training data set Learned, where D represents the total number of training samples, y (d) is the observed signal of sample d, h (d) is the real sparse signal of sample d. The training data set is input into the complex neural network to obtain the calculation result of the network forward propagation. When the total number of network layers is T, the objective function used in training is:
[0018]
[0019] in, is the parameter to be learned in the complex network, μ is the adjustment factor, and h T (y (d) ; Θ) is when the input of the network is y (d) When the output of the network layer T is (d) When the output of the neural network is T (y (d) ; Θ) and ideal output h (d) The error between them is as small as possible; the mini-batch gradient descent (MBGD) algorithm is used for network training and parameter update, and the Adam algorithm is used for optimization; the trained neural network is used for channel estimation to obtain the final estimate h.
[0020] The present invention has the following beneficial effects:
[0021] 1. The present invention provides for the first time a design method for OFDM system channel estimation based on LISTA complex network.
[0022] 2. The present invention represents the real and imaginary parts of the LISTA complex network separately as logically different real-valued entities, so that the network has richer representation capabilities.
[0023] 3. The simulation experiments verify that the network of the present invention has good estimation performance. BRIEF DESCRIPTION OF THE DRAWINGS
[0024] Figure 1 It is a schematic diagram of the flow chart of the algorithm of the present invention;
[0025] Figure 2 is a structural diagram of the LISTA complex network of the present invention;
[0026] Figure 3 This is a comparison chart of the MSE of the LISTA complex network of the present invention and other algorithms in OFDM system channel estimation;
[0027] Figure 4 It uses the LISTA complex network and other algorithms to measure the changes in the MSE with the number of iterations / network layers. DETAILED DESCRIPTION
[0028] The design method of OFDM system channel estimation based on LISTA complex network provided by the present invention mainly includes three parts: system model, channel estimation based on complex network construction and complex network parameter training. The implementation process of these three parts will be specifically introduced below. The specific steps are as follows:
[0029] First, according to the design requirements, in the OFDM system, the number of subcarriers is given as N = 512, the channel length is L = 100, and the number of pilots is P = 64, which are located at subcarriers q1, q2, ...q64 The transmission symbol of the transmitting antenna on the pilot subcarrier is x, which is a (64, 1)-dimensional vector. The pilot observation vector at the receiving end is expressed as:
[0030] y=diag(x)F L h+n (6)
[0031] In the formula, is the received signal of the receiving antenna at the pilot position, is the multipath channel impulse response vector from the base station transmitting antenna to the user, h follows the Bernoulli Gaussian distribution, and each element in h is set to 0 with a probability of 0.1 and to a sample value in the standard normal distribution with a probability of 0.9. L It is composed of the first L columns in the Fourier matrix F and the P rows corresponding to the pilot positions, with a dimension of (64, 100). The discrete Fourier transform matrix F of N points is expressed as:
[0032]
[0033] Among them, the qth row and the lth column Represents the additive Gaussian white noise of the channel. The value of each element in the noise n follows a Gaussian distribution with a mean of 0, and the variance depends on the average power of the received signal. Let A = diag(x)F L , then the channel estimation problem model (1) can be transformed into the solution of the minimum l0 norm problem, expressed as follows:
[0034]
[0035] Since the receiving vector y, observation matrix A, channel vector h and noise vector n in the OFDM system are all in complex form, most of the current deep learning architectures are based on real-valued operations and representations. Deep learning cannot directly process complex numbers. This design uses the virtual-real separation method under the complex network to process the real and imaginary parts separately, and then combine the processed real and imaginary parts to represent the complex number.
[0036] Second, expand the iterative threshold algorithm into a recurrent neural network and build a complex network LISTA with training parameters. The network structure of each layer is the same, and the total number of layers of the network is T = 20. In the formula, each complex signal has a real component and an imaginary component. The real and imaginary parts of the complex number are separated and represented as logically different real-valued entities, and the real-valued algorithm is used internally to simulate complex arithmetic, resulting in the following iterative process:
[0037]
[0038] in and and B rand B i ,y r and i , A r and A i Represents z t 、h t , B, y, the real and imaginary components of A, Represents vector h t The L1 norm of B r , B i , α t (t=0,1,...2T-1) are the parameters to be learned in the complex network, and the threshold shrinkage parameters [η st (z t ;λ)] j =sgn(z t [j])max{|z t [j]|-λ,0}(t=0,1,2,...) represents the element-wise soft threshold shrinkage function, and is the input vector of the t-th layer network, h t+1 is the output vector of the t-th layer network. The initial value h0 is a zero vector. The h of the last layer output is t This is the channel estimation value h, that is, h = h T .
[0039] Third, use the training data set to train the parameters of the complex network. When training the LISTA complex network, learn the parameters B t The value of each layer and the shrinkage function threshold λ t All through the training data set Learned, where D represents the total number of training samples, y (d) is the observed signal of sample d, h (d) is the real sparse signal of sample d. The training set and test set contain 200,000 samples and 300,000 samples respectively (50,000 samples for each signal-to-noise ratio). The training data set is input into the complex neural network to obtain the calculation result of the network forward propagation. When the total number of network layers is T, the objective function used in training is:
[0040]
[0041] in, is the parameter to be learned in the complex network, the adjustment factor μ = 0.01, and the output h of the neural network T (y (d) ; Θ) and ideal output h (d)The error between them is as small as possible; the mini-batch gradient descent (MBGD) algorithm is used for network training and parameter update, and the Adam algorithm is used for optimization. The trained neural network is used for channel estimation to obtain the final estimate h.
[0042] In order to verify the performance of the algorithm, a computer simulation was performed on the method. A 20-layer neural network was built using PyTorch, with an input layer data set size of 1000 mini batches of 200, and the learning rate of the Adam optimizer was set to 0.003 for training and parameter optimization. Figure 3 The MSE comparison chart of LISTA complex network and other algorithms in OFDM system channel estimation is shown. Compared with ISTA algorithm and FISTA algorithm, the algorithm based on LISTA complex network has better noise resistance, better reconstruction performance and more accurate estimation accuracy. In the figure, the MSE curves of ISTA and FISTA algorithms are almost the same, but the number of iterations required by FISTA is much less than that of ISTA. In order to prove the convergence of LISTA complex network, Figure 4 The paper shows how the MSE of the algorithm based on the LISTA complex network and the ISTA and FISTA algorithms change with the number of iterations when SNR=40dB. The simulation results show that the LISTA complex network can converge at around T=10 layers, and when the MSE is constant, the LISTA complex network requires fewer network layers, effectively saving network resources.
Claims
1. A design method for OFDM system channel estimation based on LISTA complex network, characterized by The complex network is used for channel reconstruction, which makes full use of the relationship between the real and imaginary parts. Compared with the real-valued algorithm, it has better generalization error and can achieve nearly optimal channel estimation. The method is carried out in the following steps: First, according to the design requirements, in the OFDM system, the number of subcarriers is N, the channel length is L, and the number of pilots is P, which are located at subcarriers q1, q2, ...q p The transmission symbol of the transmitting antenna on the pilot subcarrier is Then the pilot observation vector at the receiving end can be expressed as: y=diag(x)F L h+n (1) In the formula, is the received signal of the receiving antenna at the pilot position, is the impulse response vector of the multipath channel from the base station transmitting antenna to the user. The channel length is L, the number of non-zero positions is K, and K<<L is satisfied. That is, the time domain channel presents sparse characteristics, and most of the energy is concentrated on a small number of channel taps. F is the discrete Fourier transform matrix of N points. is a submatrix consisting of the first L columns of the Fourier transform matrix F and the P rows corresponding to the pilot positions, Denotes the additive white Gaussian noise of the channel, let A = diag(x)F L , then the channel estimation problem model (1) can be transformed into the solution of the minimum l0 norm problem, expressed as follows: Since the receiving vector y, the observation matrix A, the channel vector h and the noise vector n in the OFDM system are all in complex form, most of the current deep learning architectures are based on real-valued operations and representations, and deep learning cannot directly process complex numbers. This design uses the virtual-real separation method under the complex network to process the real part and the imaginary part respectively, and then combine the processed real and imaginary parts to represent the complex number; Second, the traditional iterative threshold algorithm process is where z t is the input vector of the tth iteration, β is the step size of the gradient update, and h t+1 is the output vector of the tth iteration, [η st (z t ;λ)] j =sgn(z t [j])max{|z t [j]|-λ,0}(t=0,1,2,…) represents the element-by-element soft threshold shrinkage function, r j Represents vector z t The jth element in h is updated using the current gradient in each iteration. t Get z t ; Third, expand the iterative threshold algorithm into a recurrent neural network and build a complex network LISTA with training parameters. The total number of layers of the network is recorded as T. In the formula, each complex signal has a real component and an imaginary component. The real and imaginary parts of the complex number are separated and represented as logically different real-valued entities. The real-valued algorithm is used internally to simulate complex arithmetic. Let B = βA T , and the following iterative process is obtained: in and and B r and B i ,y r and i , A r and A i Represents z t 、h t , B, y, the real and imaginary components of A, Represents vector h t The L1 norm of B r , B i , α t (t=0,1,…2T-1) are the parameters that need to be learned in the complex network, is the threshold shrinkage parameter; and is the input vector of the t-th layer network, h t+1 is the output vector of the t-th layer network. The initial value h0 is a zero vector. The h of the last layer output is t This is the channel estimation value h, that is, h = h T ; Fourth, use the training data set to train the parameters of the complex network; the main idea is to train a neural network with a specific architecture and fixed depth to produce the best approximation of the sparse code and learn the parameters B t The value of each layer and the shrinkage function threshold λ t All through the training data set Learned, where D represents the total number of training samples, y (d) is the observed signal of sample d, h (d) is the real sparse signal of sample d. The training data set is input into the complex neural network to obtain the calculation result of the network forward propagation. When the total number of network layers is T, the objective function used in training is: in, is the parameter to be learned in the complex network, μ is the adjustment factor, and h T (y (d) ; Θ) is when the input of the network is y (d) When the output of the network layer T is (d) When the output of the neural network is T (y (d) ; Θ) and ideal output h (d) The error between them is as small as possible; the mini-batch gradient descent (MBGD) algorithm is used for network training and parameter update, and the Adam algorithm is used for optimization; the trained neural network is used for channel estimation to obtain the final estimate h.