OTFS system channel estimation method based on ResUNet network

Abstract

The invention discloses an OTFS system channel estimation method based on a ResUNet network. The method comprises the steps that information to be transmitted is modulated through 16-QAM; inserting a pilot symbol into a data symbol to be sent to obtain a DD domain signal; performing OTFS modulation to obtain an OTFS signal; the OTFS signal reaches a receiving end through a wireless channel; converting the received signal back to a received signal of the DD domain; the method comprises the following steps: performing preliminary estimation on a DD domain signal received by a receiving end by utilizing pilot frequency information in the DD domain signal to obtain a channel matrix at a pilot frequency; respectively extracting a real part and an imaginary part of the channel matrix at the pilot frequency; the small-size channel matrix is amplified to a target matrix to serve as input of the improved ResUNet network, and a final channel matrix is obtained; performing signal detection on the original data and the noise data received by the receiving end; and carrying out 16-QAM demodulation on the detected target signal. According to the method, the pilot frequency is designed and inserted into the communication signal, deep learning is carried out on the basis of the improved ResUNet network by using the pilot frequency at the receiving end, and high-precision channel estimation is realized.

Description

Technical Field

[0001] The invention belongs to the field of channel estimation, relates to an OTFS system channel estimation technology, and specifically relates to an OTFS system channel estimation method based on a ResUNet network. Background Art

[0002] In recent years, with the rapid development of mobile communication technology, the performance of OFDM (orthogonal frequency division multiplexing) modulation technology will drop sharply in high mobility scenarios. Since OFDM cannot effectively deal with the signal frequency dispersion problem caused by time-varying channels, Orthogonal Time Frequency Space (OTFS) came into being. The key idea of ​​OTFS modulation technology is to transmit information symbols in the delay-doppler domain (DD) to solve the problem of Doppler high frequency shift in time-varying channels. In the field of channel estimation, OTFS has the significance of in-depth research.

[0003] Traditional channel estimation methods often perform poorly in complex environments such as time-varying channels and noise interference. In recent years, the Compressed Sensing (CS) theory proposed by Donoho, including convex optimization algorithms and greedy algorithms, can achieve better estimation performance on the basis of reducing pilot overhead. However, in actual wireless communication systems, it is often difficult to obtain accurate channel models and CSI, which inevitably leads to errors in channel estimation.

[0004] At present, the application of deep learning in channel estimation has become an important direction in the field of wireless communications. Traditional channel estimation methods often rely on known channel models and assumptions, while deep learning can provide more accurate estimates in complex and changing channel environments through a data-driven approach. Technologies such as deep neural networks (DNNs), convolutional neural networks (CNNs), and recurrent neural networks (RNNs) can automatically extract features from received signals, capture the nonlinear and time-varying characteristics of channels, and thus improve the accuracy of channel estimation. Compared with traditional methods, deep learning methods do not require accurate channel models, can handle complex situations such as multipath effects and signal attenuation, and have strong robustness. In particular, deep learning has shown great potential in scenarios such as massive MIMO systems, high-speed mobile communications, and OTFS systems. Although deep learning has performed well in channel estimation, it still faces challenges in terms of its huge data requirements and high computational complexity. Therefore, how to optimize deep learning models to meet the actual needs of communication systems remains a difficult problem.

[0005] On the one hand, in actual wireless communications, the received signal is usually mixed with noise and other interfering signals, which will reduce the accuracy of channel estimation. Especially under low signal-to-noise ratio conditions, the impact of noise is more significant, resulting in instability and increased errors in the estimation results; on the other hand, wireless channels are often affected by factors such as multipath propagation, frequency selective fading, and user mobility, causing channel characteristics to change over time. In high mobility scenarios, the channel state will change rapidly, making the estimation based on previous channel information inaccurate. Therefore, in the case where CSI cannot be accurately obtained, a new algorithm is urgently needed to solve the channel estimation problem in the above scenarios, thereby improving the overall performance of the OTFS system in complex communication environments. Summary of the invention

[0006] Purpose of the invention: In order to overcome the deficiencies in the prior art, a channel estimation method for an OTFS system based on a ResUNet network is provided to solve the problems of inter-subcarrier crosstalk and inter-symbol interference in the communication system and improve the reliability of the system. The present invention designs a pilot and inserts it into the communication signal. The pilot is used at the receiving end to perform deep learning based on the improved ResUNet network to achieve high-precision channel estimation; reliable communication is achieved through channel equalization.

[0007] Technical solution: To achieve the above purpose, the present invention provides an OTFS system channel estimation method based on a ResUNet network, comprising the following steps:

[0008] S1: At the transmitting end, the information to be transmitted is modulated by 16-QAM to obtain the data symbols to be sent;

[0009] S2: inserting pilot symbols into the data symbols to be sent to obtain DD domain signals;

[0010] S3: Perform OTFS modulation on the DD domain signal to obtain an OTFS signal;

[0011] S4: OTFS signal reaches the receiving end through the wireless channel;

[0012] S5: The receiving end transforms the received signal back to the received signal in the DD domain;

[0013] S6: Perform a preliminary estimation on the DD domain signal received by the receiving end using the pilot information therein to obtain a channel matrix at the pilot;

[0014] S7: extracting the real part and the imaginary part of the channel matrix at the pilot respectively as a small-size channel matrix;

[0015] S8: Enlarge the small-size channel matrix to the target matrix as the input of the improved ResUNet network. Through multiple convolutional layers and residual connections, extract the deep features in the channel, and fuse the information at different levels through skip connections to obtain the final channel matrix.

[0016] S9: Perform signal detection on the original data and noise data received by the receiving end according to the estimated channel state information, and restore the original signal;

[0017] S10: Perform 16-QAM demodulation on the detected target signal to obtain initial bit stream data.

[0018] Furthermore, the pilot symbols of the data symbols in the M×N DD domain signal grids in step S2 are used to obtain a signal in the delay Doppler domain, which is expressed as: x[k,l], k=0,1,2,...,N-2,N-1, l=0,1,2,...,M-2,M-1; the pilot symbol is expressed as Insert pilot symbols, the pilot symbol range is

[0019] OTFS frame transmission is represented as:

[0020]

[0021] Among them, x[k,l] is the DD domain signal.

[0022] Furthermore, in step S3, the DD domain signal x[k, l] is converted to the TF domain through ISFFT to obtain a TF domain signal, which is expressed as:

[0023]

[0024] Wherein, X[n,m] represents the TF time-frequency domain signal, M and N represent the size of the delay and Doppler dimensions respectively, n and m are time-frequency domain indexes, n = 0, 1, 2, K, N-2, N-1, m = 0, 1, 2, K, M-2, M-1, x[k,l] represents the signal in the delay-Doppler domain, k and l represent the indexes in the delay domain and Doppler domain respectively, j is an imaginary unit, e represents the exponential function, and π is a constant;

[0025] After obtaining the TF time-frequency domain signal X[n,m], the Heisenberg transform is used to convert X[n,m] into a continuous time-domain transmission signal s(t), namely the OTFS signal, which is expressed as:

[0026]

[0027] Where T is the symbol duration, g tx (·) represents the shaping filter function of the transmitted pulse, and Δf represents the frequency interval.

[0028] Furthermore, in step S4, the time domain transmission signal s(t) reaches the receiving end after passing through the wireless channel, and the receiving end receives the following signal r(t):

[0029] r(t)=∫ v ∫ τ h(τ,v)s(tv)e j2πv(t-v) dτdv+n(t)

[0030] in, n(t) represents noise, h(τ,v) is the DD domain channel response, which reflects the propagation of the signal in the delay-Doppler domain; there are a total of P paths in the channel, and the channel amplitude response of the Pth path is β P ; In the delay domain and Doppler domain, the impulse function δ(τ-τ i ) and δ(vv i ) is used to represent the concentrated position of the signal; τ, v represent the time delay and Doppler frequency shift respectively;

[0031] Delay τ i and Doppler shift ν i The characteristics corresponding to the i-th path are expressed as:

[0032]

[0033] Among them, τ i Used to represent the delay of the i-th path, ν i The integer part represents the integer Doppler shift, and the fractional part reflects the offset from the most recent Doppler shift.

[0034] Furthermore, the step S5 specifically includes:

[0035] A1: Compare the received signal r(t) with the receiving end pulse shaping waveform g rx (t) Perform matched filtering to obtain the received data, and use the following formula to calculate the received signal in the time-frequency domain:

[0036]

[0037] in, represents the cross fuzzy function, represents the matched filter function, t′ represents the integral variable, and represents the time offset;

[0038] A2: After the OTFS signal passes through the wireless channel, it is demodulated by OTFS to obtain the OTFS symbol frame in the DD domain;

[0039] A3: Sample Y(t,f) at intervals of t=nT,f=mΔf to obtain discrete time-frequency domain received signals. This step is called Wigner transform, which is expressed as:

[0040]

[0041] A4: Use SFFT to process the TF domain received signal Y[n,m] to obtain the DD domain received signal y[k,l], which is expressed as:

[0042]

[0043] Due to the unidirectionality of time and the bidirectionality of relative speed, it can be expressed as:

[0044] c=λ·f

[0045]

[0046] Where c is the propagation speed, λ is the wavelength, and f is the carrier frequency. Since the angle changes, that is, cosθ can be ±1, the DD domain signal has a delay dimension of 0 to τ. max One-way extension, in the Doppler dimension -v max ~v max Bidirectional extension, we can see that if the sender (l p ,k p ) is regarded as the reference signal. After the data frame is transmitted through the channel, the range of dispersion at the receiving end is The data in the pilot extension signal area at the receiving end is expressed as:

[0047]

[0048] Among them, for It can be seen that if there exists a Doppler tap kk p and delay tap ll p The path, that is The above formula is expressed as:

[0049]

[0050] Otherwise, y[k,l]=v[k,l];v[k,l]~CN(0,σ 2 ) is a variable with variance σ 2 Additive white Gaussian noise.

[0051] Furthermore, the step S6 specifically includes:

[0052] The transmitted pilot symbol is represented as The received signal can be expressed as y[k,l], the channel matrix at the pilot is H, and the channel noise is v[k,l]. The relationship is expressed as:

[0053]

[0054] Furthermore, in step S7, the real part H of the channel matrix H at the pilot is extracted respectively. real and the imaginary part H imag , these two matrices are used as the data of the two channels respectively, which are regarded as small-size channel matrices and used as the input of the ResUNet network. The relationship is expressed as follows:

[0055]

[0056] Furthermore, the step S8 specifically includes:

[0057] B1: Insert zeros between every two adjacent pixels in the input channel matrix H, perform convolution operation on the channel matrix after inserting zero values, and finally output the target matrix H′; expand the channel matrix H by inserting zero values ​​and performing convolution. The process is expressed as:

[0058] H′=H*K

[0059] Among them, the convolution kernel is K, and * represents the convolution operation;

[0060] B2: The ResUNet network consists of an encoder and a decoder;

[0061] The encoder is divided into five stages, each of which consists of a residual block consisting of two convolutional layers and a residual connection replacing the original convolutional block;

[0062] The encoder part, with the input as the enlarged target matrix H′, of size M×N×2, runs:

[0063] In the first stage, both convolutional layers use 64 3×3 convolution kernels to extract low-level features of the input matrix;

[0064] In the second stage, both convolutional layers use 128 3×3 convolution kernels to further extract deeper features;

[0065] In the third stage, both convolutional layers use 256 3×3 convolution kernels to further extract deeper features;

[0066] In the fourth stage, both convolutional layers use 512 3×3 convolution kernels to further extract deeper features;

[0067] In the fifth stage, the two convolutional layers still use 512 3×3 convolution kernels and maintain the original number of channels;

[0068] A ReLU activation function is applied after each convolutional layer to introduce nonlinear features;

[0069] A residual connection is added after every two convolution operations and ReLU activation functions to solve the gradient vanishing problem, making the model easier to converge and reducing the difficulty of model training to enhance the feature extraction capability;

[0070] After each stage, a 2×2 convolution kernel is used for downsampling to help reduce the size of the channel matrix, thereby reducing the amount of computation and retaining important features;

[0071] The decoder is divided into four stages. Each convolution block in each stage is still replaced by a residual block, and the upsampled feature map is concatenated with the corresponding encoder feature received through the skip connection, and feature fusion is performed through the residual block.

[0072] The operation of the decoder part includes:

[0073] In the first stage, both convolutional layers use 512 3×3 convolution kernels to restore the channel matrix features;

[0074] In the second stage, both convolutional layers use 256 3×3 convolution kernels to further restore the channel matrix features;

[0075] In the third stage, both convolutional layers use 128 3×3 convolution kernels to further restore the channel matrix features;

[0076] In the fourth stage, both convolutional layers use 64 3×3 convolution kernels to further restore the channel matrix features;

[0077] At each stage, the channel matrix size is first increased by upsampling. Here, bilinear interpolation is used instead of transposed convolution for upsampling. There is no need to change the height and width of the input feature layer, which avoids possible artifacts in transposed convolution and ensures the integrity of the information.

[0078] B3: Use skip connections to concatenate the upsampled channel feature matrix with the corresponding encoder feature matrix, so that the network can fuse information at different levels;

[0079] A ReLU activation function is applied after each convolutional layer to introduce nonlinear features;

[0080] A residual connection is added after every two convolution operations and ReLU activation functions to solve the gradient vanishing problem, making the model easier to converge and reducing the difficulty of model training to enhance the feature extraction capability;

[0081] By gradually recovering the information, two 1×1 convolution kernels are finally used to restore the original number of channels, generate a channel feature matrix, and obtain an accurate channel estimation result.

[0082] Furthermore, the residual block in step B3 has two paths for output. The first is the output after two 3×3 convolutions and the nonlinear transformation of the ReLU activation function. At this time, the output F(H′) represents the residual of H′ after the convolution and nonlinear transformation layers. The second is mapped to the output through the upper jump connection. The convolution formula is expressed as:

[0083]

[0084] Among them, New-Matrix represents the matrix size after convolution, Matrix represents the matrix size before convolution, Kernelsize represents the convolution kernel size, Padding represents zero padding, and stride represents the step size; the first convolution operation of the input layer sets the convolution kernel size to 3, zero padding to 1, and stride to 1, so the size after convolution remains unchanged, and the number of channels increases to 64. The second convolution operation of the input layer sets the convolution kernel size to 3, zero padding to 1, and stride to 1, so the size after convolution remains unchanged, and the number of channels remains unchanged;

[0085] The ReLU function has a huge acceleration effect on the convergence of stochastic gradient descent, and to a certain extent alleviates the gradient vanishing problem of neural networks; the ReLU function formula is expressed as:

[0086] f(x)=max(0,x)

[0087] The output y in the residual block is expressed as:

[0088] y=F(H′)+H′

[0089] Adding residual connections can solve the gradient vanishing problem, making the model easier to converge and reducing the difficulty of model training. During the back propagation process, the gradient is transmitted through the chain rule. In ordinary networks, if the number of network layers is large, the gradient will gradually decrease or even approach 0 during back propagation, that is, the gradient disappears. With the help of residual connections, assuming that the loss function is L and the output is y, the gradient of the calculated loss relative to the input H′ is expressed as:

[0090]

[0091] Substituting into the above formula we get:

[0092]

[0093] Due to the existence of the identity mapping, even if Very small, resulting in gradient attenuation, but the gradient will not disappear completely;

[0094] During the training process, the model parameters randomly generate M training samples, (H i ,G i) is the i-th group of training samples; H i For the channel matrix obtained by preliminary estimation, the imaginary and real parts of the estimated matrix are divided into two channels of a channel matrix, that is, Then there is M×N is the size of the channel matrix; G i is the corresponding true channel matrix; the network is trained using the loss function λ 2 It is expressed as:

[0095]

[0096] Where T is the size of the training set; f θ (H i ) is the channel matrix output by the model; the loss value of the loss function reflects the good convergence speed of the improved U-net structure;

[0097] The mean square error (MSE) is used during network training to measure the error between the network output and the actual value, which can be expressed as:

[0098]

[0099] In order to obtain the best set of network parameters and make the training results gradually approach the optimal value, the Adam optimizer is used, the adaptive learning rate is set, and different learning rates are applied to each parameter to make the optimization process more stable.

[0100] Furthermore, the step S9 specifically includes:

[0101] Subtract the pilot signal from the received signal y to obtain y d , which will be used to detect the expression:

[0102]

[0103] The purpose here is to d and Estimated data symbol vector x d ;

[0104] initialization: Prepare initial conditions for algorithm iterations;

[0105] Start the loop:

[0106] Calculate the variance of the propagation model: in is the transfer matrix, represents the square of its modulus, is the iterated noise variance;

[0107] Compute the current signal forecast:

[0108] Update the variance of the signal: v s =1. / (v p +ε -1 1);

[0109] Calculate the current noise term: s t =v s ·(rp);

[0110] Update the variance of the received signal:

[0111] Update the estimate of the received signal:

[0112] Compute the posterior distribution for each possible symbol:

[0113] Normalize the posterior distribution so that the probability of each symbol sums to 1:

[0114] Compute the next signal estimate:

[0115] Update the noise variance of the signal:

[0116] Update the number of iterations: t = t + 1;

[0117] Until terminated.

[0118] Beneficial effects: Compared with the prior art, the present invention designs a pilot signal and inserts it into the communication signal, so that the channel state can be obtained more accurately at the receiving end, thereby effectively reducing interference. The improved ResUNet network can perform deep learning based on the pilot signal, extract the complex characteristics of the channel from the pilot signal through adaptive learning, and estimate the channel state information very accurately; finally, the channel equalization technology is used to compensate for the distortion caused by the channel, and reliable communication is achieved, which solves the problems of crosstalk between subcarriers and inter-symbol interference in the communication system and improves the reliability of the OTFS system. BRIEF DESCRIPTION OF THE DRAWINGS

[0119] Figure 1 is a flow chart of the method of the present invention;

[0120] Figure 2 Insertion pattern diagram for pilots;

[0121] Figure 3 This is the ResUNet network architecture diagram;

[0122] Figure 4 This is the internal residual block structure diagram of the ResUNet network. DETAILED DESCRIPTION

[0123] The present invention is further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, various equivalent forms of modifications to the present invention by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0124] like Figure 1 As shown, the present invention provides an OTFS system channel estimation method based on a ResUNet network, comprising the following steps:

[0125] S1: At the transmitting end, the information to be transmitted is modulated by 16-QAM to obtain the data symbols to be sent;

[0126] S2: inserting pilot symbols into the data symbols to be sent to obtain DD domain signals;

[0127] like Figure 2 As shown, the M×N DD domain signal grids contain pilot symbols of data symbols, and the delay Doppler domain signal is obtained, which is expressed as: x[k,l], k = 0, 1, 2, ..., N-2, N-1, l = 0, 1, 2, ..., M-2, M-1; the pilot symbol is expressed as Insert pilot symbols, the pilot symbol range is

[0128] OTFS frame transmission is represented as:

[0129]

[0130] Among them, x[k,l] is the DD domain signal.

[0131] S3: Perform OTFS modulation on the DD domain signal to obtain an OTFS signal;

[0132] The DD domain signal x[k,l] is converted to the TF domain through ISFFT to obtain the TF domain signal, which is expressed as:

[0133]

[0134] Wherein, X[n,m] represents the TF time-frequency domain signal, M and N represent the size of the delay and Doppler dimensions respectively, n and m are time-frequency domain indexes, n = 0, 1, 2, K, N-2, N-1, m = 0, 1, 2, K, M-2, M-1, x[k,l] represents the signal in the delay-Doppler domain, k and l represent the indexes in the delay domain and Doppler domain respectively, j is an imaginary unit, e represents the exponential function, and π is a constant;

[0135] After obtaining the TF time-frequency domain signal X[n,m], the Heisenberg transform is used to convert X[n,m] into a continuous time-domain transmission signal s(t), namely the OTFS signal, which is expressed as:

[0136]

[0137] Where T is the symbol duration, g tx (·) represents the shaping filter function of the transmitted pulse, and Δf represents the frequency interval.

[0138] S4: OTFS signal reaches the receiving end through the wireless channel;

[0139] The time domain transmitted signal s(t) reaches the receiving end after passing through the wireless channel, and the receiving end receives the following signal r(t):

[0140] r(t)=∫ v ∫ τ h(τ,v)s(tv)e j2πv(t-v) dτdv+n(t)

[0141] in, n(t) represents noise, h(τ,ν) is the DD domain channel response, which reflects the propagation of the signal in the delay-Doppler domain; there are a total of P paths in the channel, and the channel amplitude response of the Pth path is β P ; In the delay domain and Doppler domain, the impulse function δ(τ-τ i ) and δ(ν-ν i ) represents the concentrated position of the signal; τ, ν represent the time delay and Doppler frequency shift respectively;

[0142] Delay τ i and Doppler shift ν i The characteristics corresponding to the i-th path are expressed as:

[0143]

[0144] Among them, τ i Used to represent the delay of the i-th path, v i The integer part represents the integer Doppler shift, and the fractional part reflects the offset from the most recent Doppler shift.

[0145] S5: The receiving end transforms the received signal back to the received signal in the DD domain, which specifically includes:

[0146] A1: Compare the received signal r(t) with the receiving end pulse shaping waveform g rx (t) Perform matched filtering to obtain the received data, and use the following formula to calculate the received signal in the time-frequency domain:

[0147]

[0148] in, represents the cross fuzzy function, represents the matched filter function, t′ represents the integral variable, i.e., the time offset;

[0149] A2: After the OTFS signal passes through the wireless channel, it is demodulated by OTFS to obtain the OTFS symbol frame in the DD domain;

[0150] A3: Sample Y(t,f) at intervals of t=nT,f=mΔf to obtain discrete time-frequency domain received signals. This step is called Wigner transform, which is expressed as:

[0151]

[0152] A4: Use SFFT to process the TF domain received signal Y[n,m] to obtain the DD domain received signal y[k,l], which is expressed as:

[0153]

[0154] Due to the unidirectionality of time and the bidirectionality of relative speed, it can be expressed as:

[0155]

[0156] Where c is the propagation speed, λ is the wavelength, and f is the carrier frequency. Since the angle changes, that is, cosθ can be ±1, the DD domain signal has a delay dimension of 0 to τ. max One-way extension, in the Doppler dimension -v max ~v max Bidirectional extension, we can see that if the sender (l p ,k p ) is regarded as the reference signal. After the data frame is transmitted through the channel, the range of dispersion at the receiving end is The data in the pilot extension signal area at the receiving end is expressed as:

[0157]

[0158] Among them, for It can be seen that if there exists a Doppler tap kk p and delay tap ll p The path, that is The above formula is expressed as:

[0159]

[0160] Otherwise, y[k,l]=v[k,l];v[k,l]~CN(0,σ2 ) is a variable with variance σ 2 Additive white Gaussian noise.

[0161] S6: Perform a preliminary estimation on the DD domain signal received by the receiving end using the pilot information therein to obtain a channel matrix at the pilot, specifically including:

[0162] The transmitted pilot symbol is represented as The received signal can be expressed as y[k,l], the channel matrix at the pilot is H, and the channel noise is v[k,l]. The relationship is expressed as:

[0163]

[0164] S7: Extract the real part H of the channel matrix H at the pilot signal respectively real and the imaginary part H imag , these two matrices are used as the data of the two channels respectively, which are regarded as small-size channel matrices and used as the input of the ResUNet network. The relationship is expressed as follows:

[0165]

[0166] S8: Enlarge the small-size channel matrix to the target matrix as the input of the improved ResUNet network. Through multiple convolutional layers and residual connections, extract the deep features in the channel, and fuse the information at different levels through skip connections to obtain the final channel matrix.

[0167] Specifically include:

[0168] B1: Insert zeros between every two adjacent pixels in the input channel matrix H, perform convolution operation on the channel matrix after inserting zero values, and finally output the target matrix H′; expand the channel matrix H by inserting zero values ​​and performing convolution. The process is expressed as:

[0169] H′=H*K

[0170] Among them, the convolution kernel is K, and * represents the convolution operation;

[0171] B2: If Figure 3 As shown, the ResUNet network consists of an encoder and a decoder;

[0172] The encoder is divided into five stages, each of which consists of a residual block consisting of two convolutional layers and a residual connection replacing the original convolutional block;

[0173] The encoder part, with the input as the enlarged target matrix H′, of size M×N×2, runs:

[0174] In the first stage, both convolutional layers use 64 3×3 convolution kernels to extract low-level features of the input matrix;

[0175] In the second stage, both convolutional layers use 128 3×3 convolution kernels to further extract deeper features;

[0176] In the third stage, both convolutional layers use 256 3×3 convolution kernels to further extract deeper features;

[0177] In the fourth stage, both convolutional layers use 512 3×3 convolution kernels to further extract deeper features;

[0178] In the fifth stage, the two convolutional layers still use 512 3×3 convolution kernels and maintain the original number of channels;

[0179] A ReLU activation function is applied after each convolutional layer to introduce nonlinear features;

[0180] A residual connection is added after every two convolution operations and ReLU activation functions to solve the gradient vanishing problem, making the model easier to converge and reducing the difficulty of model training to enhance the feature extraction capability;

[0181] After each stage, a 2×2 convolution kernel is used for downsampling to help reduce the size of the channel matrix, thereby reducing the amount of computation and retaining important features;

[0182] The decoder is divided into four stages. Each convolution block in each stage is still replaced by a residual block, and the upsampled feature map is concatenated with the corresponding encoder feature received through the skip connection, and feature fusion is performed through the residual block.

[0183] The operation of the decoder part includes:

[0184] In the first stage, both convolutional layers use 512 3×3 convolution kernels to restore the channel matrix features;

[0185] In the second stage, both convolutional layers use 256 3×3 convolution kernels to further restore the channel matrix features;

[0186] In the third stage, both convolutional layers use 128 3×3 convolution kernels to further restore the channel matrix features;

[0187] In the fourth stage, both convolutional layers use 64 3×3 convolution kernels to further restore the channel matrix features;

[0188] At each stage, the channel matrix size is first increased by upsampling. Here, bilinear interpolation is used instead of transposed convolution for upsampling. There is no need to change the height and width of the input feature layer, which avoids possible artifacts in transposed convolution and ensures the integrity of the information.

[0189] B3: Use skip connections to concatenate the upsampled channel feature matrix with the corresponding encoder feature matrix, so that the network can fuse information at different levels;

[0190] A ReLU activation function is applied after each convolutional layer to introduce nonlinear features;

[0191] A residual connection is added after every two convolution operations and ReLU activation functions to solve the gradient vanishing problem, making the model easier to converge and reducing the difficulty of model training to enhance the feature extraction capability;

[0192] By gradually recovering the information, two 1×1 convolution kernels are finally used to restore the original number of channels, generate a channel feature matrix, and obtain an accurate channel estimation result.

[0193] like Figure 4 As shown in the figure, the residual block has two output paths. The first is the output after two 3×3 convolutions and the nonlinear transformation of the ReLU activation function. At this time, the output F(H′) represents the residual of H′ after the convolution and nonlinear transformation layers. The second is mapped to the output through the upper jump connection. The convolution formula is expressed as:

[0194]

[0195] Among them, New-Matrix represents the matrix size after convolution, Matrix represents the matrix size before convolution, Kernelsize represents the convolution kernel size, Padding represents zero padding, and stride represents the step size. In this embodiment, the first convolution operation of the input layer sets the convolution kernel size to 3, the zero padding to 1, and the step size to 1, so the size after convolution remains unchanged, and the number of channels increases to 64. The second convolution operation of the input layer sets the convolution kernel size to 3, the zero padding to 1, and the step size to 1, so the size after convolution remains unchanged, and the number of channels remains unchanged;

[0196] The ReLU function has a huge acceleration effect on the convergence of stochastic gradient descent, and to a certain extent alleviates the gradient vanishing problem of neural networks; the ReLU function formula is expressed as:

[0197] f(x)=max(0,x)

[0198] The output y in the residual block is expressed as:

[0199] y=F(H′)+H′

[0200] Adding residual connections can solve the gradient vanishing problem, making the model easier to converge and reducing the difficulty of model training. During the back propagation process, the gradient is transmitted through the chain rule. In ordinary networks, if the number of network layers is large, the gradient will gradually decrease or even approach 0 during back propagation, that is, the gradient disappears. With the help of residual connections, assuming that the loss function is L and the output is y, the gradient of the calculated loss relative to the input H′ is expressed as:

[0201]

[0202] Substituting into the above formula we get:

[0203]

[0204] Due to the existence of the identity mapping, even if Very small, resulting in gradient attenuation, but the gradient will not disappear completely;

[0205] During the training process, the model parameters randomly generate M training samples, (H i ,G i ) is the i-th group of training samples; H i For the channel matrix obtained by preliminary estimation, the imaginary and real parts of the estimated matrix are divided into two channels of a channel matrix, that is, Then there is M×N is the size of the channel matrix; G i is the corresponding true channel matrix; the network is trained using the loss function λ 2 It is expressed as:

[0206]

[0207] Where T is the size of the training set; f θ (H i ) is the channel matrix output by the model; the loss value of the loss function reflects the good convergence speed of the improved U-net structure;

[0208] The mean square error (MSE) is used during network training to measure the error between the network output and the actual value, which can be expressed as:

[0209]

[0210] In order to obtain the best set of network parameters and make the training results gradually approach the optimal value, the Adam optimizer is used, the adaptive learning rate is set, and different learning rates are applied to each parameter to make the optimization process more stable.

[0211] S9: Perform signal detection on the original data and noise data received by the receiving end according to the estimated channel state information, and restore the original signal;

[0212] Specifically include:

[0213] Subtract the pilot signal from the received signal y to obtain y d , which will be used to detect the expression:

[0214]

[0215] The purpose here is to d and Estimated data symbol vector x d ;

[0216] initialization: Prepare initial conditions for algorithm iterations;

[0217] Start the loop:

[0218] Calculate the variance of the propagation model: in is the transfer matrix, represents the square of its modulus, is the iterated noise variance;

[0219] Compute the current signal forecast:

[0220] Update the variance of the signal: v s =1. / (v p +ε -1 1);

[0221] Calculate the current noise term: s t =v s ·(rp);

[0222] Update the variance of the received signal:

[0223] Update the estimate of the received signal:

[0224] Compute the posterior distribution for each possible symbol:

[0225] Normalize the posterior distribution so that the probability of each symbol sums to 1:

[0226] Compute the next signal estimate:

[0227] Update the noise variance of the signal:

[0228] Update the number of iterations: t = t + 1;

[0229] Until terminated.

[0230] S10: Perform 16-QAM demodulation on the detected target signal to obtain initial bit stream data.

Claims

1. A channel estimation method for an OTFS system based on a ResUNet network, characterized in that: The steps include: S1: At the transmitting end, the information to be transmitted is modulated by 16-QAM to obtain the data symbols to be sent; S2: inserting pilot symbols into the data symbols to be sent to obtain DD domain signals; S3: Perform OTFS modulation on the DD domain signal to obtain an OTFS signal; S4: OTFS signal reaches the receiving end through the wireless channel; S5: The receiving end transforms the received signal back to the received signal in the DD domain; S6: Perform a preliminary estimation on the DD domain signal received by the receiving end using the pilot information therein to obtain a channel matrix at the pilot; S7: extracting the real part and the imaginary part of the channel matrix at the pilot respectively as a small-size channel matrix; S8: Enlarge the small-size channel matrix to the target matrix as the input of the improved ResUNet network. Through multiple convolutional layers and residual connections, extract the deep features in the channel, and fuse the information at different levels through skip connections to obtain the final channel matrix. S9: Perform signal detection on the original data and noise data received by the receiving end according to the estimated channel state information, and restore the original signal; S10: Perform 16-QAM demodulation on the detected target signal to obtain initial bit stream data.

2. According to claim 1, a channel estimation method for an OTFS system based on a ResUNet network is characterized in that: The pilot symbols of the data symbols in the M×N DD domain signal grids in step S2 are used to obtain the delay Doppler domain signal, which is expressed as: x[k,l], k=0,1,2,...,N-2,N-1, l=0,1,2,...,M-2,M-1; the pilot symbols are expressed as Insert pilot symbols, the pilot symbol range is OTFS frame transmission is represented as: Among them, x[k,l] is the DD domain signal.

3. According to claim 2, a channel estimation method for an OTFS system based on a ResUNet network is characterized in that: In step S3, the DD domain signal x[k, l] is converted to the TF domain through ISFFT to obtain a TF domain signal, which is expressed as: Wherein, X[n,m] represents the TF time-frequency domain signal, M and N represent the size of the delay and Doppler dimensions respectively, n and m are time-frequency domain indexes, n = 0, 1, 2, K, N-2, N-1, m = 0, 1, 2, K, M-2, M-1, x[k,l] represents the signal in the delay-Doppler domain, k and l represent the indexes in the delay domain and Doppler domain respectively, j is an imaginary unit, e represents the exponential function, and π is a constant; After obtaining the TF time-frequency domain signal X[n,m], the Heisenberg transform is used to convert X[n,m] into a continuous time-domain transmission signal s(t), namely the OTFS signal, which is expressed as: Where T is the symbol duration, g tx (·) represents the shaping filter function of the transmitted pulse, and Δf represents the frequency interval.

4. According to claim 3, a channel estimation method for an OTFS system based on a ResUNet network is characterized in that: In step S4, the time domain transmission signal s(t) reaches the receiving end after passing through the wireless channel, and the receiving end receives the following signal r(t): r(t)=∫ v ∫ τ h(τ,v)s(t-v)e j2πv(t-v) dτdv+n(t) in, n(t) represents noise, h(τ,v) is the DD domain channel response, which reflects the propagation of the signal in the delay-Doppler domain; there are a total of P paths in the channel, and the channel amplitude response of the Pth path is β P ; In the delay domain and Doppler domain, the impulse function δ(τ-τ i ) and δ(vv i ) represents the concentrated position of the signal; τ, ν represent the time delay and Doppler frequency shift respectively; Delay τ i and Doppler shift v i The characteristics corresponding to the i-th path are expressed as: Among them, τ i Used to represent the delay of the i-th path, v i The integer part represents the integer Doppler shift, and the fractional part reflects the offset from the most recent Doppler shift.

5. According to claim 4, a channel estimation method for an OTFS system based on a ResUNet network is characterized in that: The step S5 specifically includes: A1: Compare the received signal r(t) with the receiving end pulse shaping waveform g rx (t) Perform matched filtering to obtain the received data, and use the following formula to calculate the received signal in the time-frequency domain: in, represents the cross fuzzy function, represents the matched filter function, t′ represents the integral variable, i.e., the time offset; A2: After the OTFS signal passes through the wireless channel, it is demodulated by OTFS to obtain the OTFS symbol frame in the DD domain; A3: Sample Y(t,f) at intervals of t=nT,f=mΔf to obtain discrete time-frequency domain received signals. This step is called Wigner transform, which is expressed as: A4: Use SFFT to process the TF domain received signal Y[n,m] to obtain the DD domain received signal y[k,l], which is expressed as: Due to the unidirectionality of time and the bidirectionality of relative speed, it can be expressed as: c=λ·f Where c is the propagation speed, λ is the wavelength, and f is the carrier frequency. Since the angle changes, that is, cosθ can be ±1, the DD domain signal has a delay dimension of 0 to τ. max One-way extension, in the Doppler dimension -v max ~v max Bidirectional extension, we can see that if the sender (l p ,k p ) is regarded as the reference signal. After the data frame is transmitted through the channel, the range of dispersion at the receiving end is The data in the pilot extension signal area at the receiving end is expressed as: Among them, for If there exists a Doppler tap kk p and delay tap ll p The path, that is The above formula is expressed as: Otherwise, y[k,l]=v[k,l];v[k,l]~CN(0,σ 2 ) is a variable with variance σ 2 Additive white Gaussian noise.

6. According to claim 5, a channel estimation method for an OTFS system based on a ResUNet network is characterized in that: The step S6 specifically includes: The transmitted pilot symbol is represented as The received signal can be expressed as y[k,l], the channel matrix at the pilot is H, and the channel noise is v[k,l]. The relationship is expressed as:

7. The OTFS system channel estimation method based on ResUNet network according to claim 1, characterized in that: In step S7, the real part H of the channel matrix H at the pilot is extracted respectively. real and the imaginary part H imag , these two matrices are used as the data of the two channels respectively, which are regarded as small-size channel matrices and used as the input of the ResUNet network. The relationship is expressed as follows:

8. The OTFS system channel estimation method based on ResUNet network according to claim 1, characterized in that: The step S8 specifically includes: B1: Insert zeros between every two adjacent pixels in the input channel matrix H, perform convolution operation on the channel matrix after inserting zero values, and finally output the target matrix H′; expand the channel matrix H by inserting zero values ​​and performing convolution. The process is expressed as: H′=H*K Among them, the convolution kernel is K, and * represents the convolution operation; B2: The ResUNet network consists of an encoder and a decoder; The encoder is divided into five stages, each of which consists of a residual block consisting of two convolutional layers and a residual connection replacing the original convolutional block; The encoder part, with the input as the enlarged target matrix H′, of size M×N×2, runs: In the first stage, both convolutional layers use 64 3×3 convolution kernels to extract low-level features of the input matrix; In the second stage, both convolutional layers use 128 3×3 convolution kernels to further extract deeper features; In the third stage, both convolutional layers use 256 3×3 convolution kernels to further extract deeper features; In the fourth stage, both convolutional layers use 512 3×3 convolution kernels to further extract deeper features; In the fifth stage, the two convolutional layers still use 512 3×3 convolution kernels and maintain the original number of channels; A ReLU activation function is applied after each convolutional layer to introduce nonlinear features; Add a residual connection after every two convolution operations and ReLU activation functions; After each stage, a 2×2 convolution kernel is used for downsampling to help reduce the size of the channel matrix; The decoder is divided into four stages. Each convolutional block in each stage is replaced by a residual block, and the upsampled feature map is concatenated with the corresponding encoder feature received through the skip connection, and feature fusion is performed through the residual block; The operation of the decoder part includes: In the first stage, both convolutional layers use 512 3×3 convolution kernels to restore the channel matrix features; In the second stage, both convolutional layers use 256 3×3 convolution kernels to further restore the channel matrix features; In the third stage, both convolutional layers use 128 3×3 convolution kernels to further restore the channel matrix features; In the fourth stage, both convolutional layers use 64 3×3 convolution kernels to further restore the channel matrix features; In each stage, the channel matrix size is first increased by upsampling, and bilinear interpolation is used instead of transposed convolution for upsampling; B3: Use skip connections to concatenate the upsampled channel feature matrix with the corresponding encoder feature matrix, so that the network can fuse information at different levels; A ReLU activation function is applied after each convolutional layer to introduce nonlinear features; Add a residual connection after every two convolution operations and ReLU activation functions; By gradually recovering the information, two 1×1 convolution kernels are finally used to restore the original number of channels, generate a channel feature matrix, and obtain an accurate channel estimation result.

9. The OTFS system channel estimation method based on ResUNet network according to claim 8, characterized in that: In step B3, the residual block has two output paths. The first is the output after two 3×3 convolutions and the nonlinear transformation of the ReLU activation function. At this time, the output F(H′) represents the residual of H′ after the convolution and nonlinear transformation layers. The second is mapped to the output through the upper jump connection. The convolution formula is expressed as: Among them, New-Matrix represents the matrix size after convolution, Matrix represents the matrix size before convolution, Kernelsize represents the convolution kernel size, Padding represents zero padding, and stride represents the step size; The ReLU function formula is expressed as: f(x)=max(0,x) The output y in the residual block is expressed as: y=F(H′)+H′ With the help of residual connections, assuming the loss function is L and the output is y, the gradient of the loss with respect to the input H′ is expressed as: Substituting into the above formula we get: Due to the existence of the identity mapping, even if Very small, resulting in gradient attenuation, but the gradient will not disappear completely; During the training process, the model parameters randomly generate M training samples, (H i ,G i ) is the i-th group of training samples; H i For the channel matrix obtained by preliminary estimation, the imaginary and real parts of the estimated matrix are divided into two channels of a channel matrix, that is, Then there is M×N is the size of the channel matrix; G i is the corresponding true channel matrix; the network is trained using the loss function λ 2 It is expressed as: Where T is the size of the training set; f θ (H i ) is the channel matrix output by the model.

10. The OTFS system channel estimation method based on ResUNet network according to claim 1, characterized in that: The step S9 specifically includes: Subtract the pilot signal from the received signal y to obtain y d , which will be used to detect the expression: From y d and Estimated data symbol vector x d ; Initialization: s (-1) =0, t = 0 prepares the initial conditions for the iteration of the algorithm; Start the loop: Calculate the variance of the propagation model: in is the transfer matrix, represents the square of its modulus, is the iterated noise variance; Compute the current signal forecast: Update the variance of the signal: v s =1. / (v p +ε -1 1); Calculate the current noise term: s t =v s ·(rp); Update the variance of the received signal: Update the estimate of the received signal: Compute the posterior distribution for each possible symbol: Normalize the posterior distribution so that the probability of each symbol sums to 1: Compute the next signal estimate: Update the noise variance of the signal: Update the number of iterations: t = t + 1; Until terminated.

Images