DNN-based OTFS time-frequency domain situation awareness and channel estimation method

Abstract

A DNN-based OTFS time-frequency domain situation awareness and channel estimation method comprises the following steps: modeling a frequency domain selective fading channel in a wireless communication process into an autoregression process, and further converting an estimation problem into an autoregression coefficient estimation process; the method comprises the following steps: acquiring training data by performing discrete sampling, DFT (Discrete Fourier Transform), characteristic matrix representation, channel prior information generation and other processes on a channel impulse response function, finishing a preprocessing process of the training data, designing a neural network based on data and model driving, training the neural network through a large amount of prior channel data, and obtaining the channel impulse response function. The optimal frequency domain related parameter estimation under the minimum mean square error of the network is solved through offline training and online estimation, then corresponding channel response is obtained, and the algorithm can achieve lower bit error rate and complexity under the condition of the same signal-to-noise ratio.

Description

Technical Field

[0001] The present invention belongs to the field of wireless communication, and specifically relates to a DNN-based OTFS time-frequency domain situation awareness and channel estimation method, which can be applicable to the development of future wireless communication technology and the communication transmission guarantee for high-dynamic targets. Background Art

[0002] With the development of wireless communication technology, the demand for more efficient, reliable and higher service quality communications has promoted the development of orthogonal frequency division multiplexing, which has achieved high-speed, low-latency wireless communication to a certain extent, and has good anti-multipath interference capabilities to ensure reliable data transmission. However, in high-speed mobile scenarios, OFDM technology will not only reduce the system's bit error rate, but also increase the complexity of the receiver due to the influence of multipath effects. Therefore, based on the communication needs of 6G, the Orthogonal Time Frequency Space (OTFS) technology is proposed, which can achieve high robustness communication in high multipath and Doppler effect environments, can significantly reduce signal distortion and inter-symbol interference, and reduce the complexity of the communication system, with superior performance-complexity trade-offs.

[0003] However, OTFS technology still faces many problems that need to be solved. The channel estimation and pilot design of the OTFS system are more challenging. For example, the pilot overhead is high, the signal-to-noise ratio requirement of the pilot signal is high, and the computational complexity is high. The sparse characteristics of the delay-Doppler domain channel and the time-correlated characteristics of the channel have not been fully utilized in the current channel estimation scheme. In addition, the traditional OTFS channel estimation method has low spectral efficiency and cannot be applied to communication scenarios with high Doppler frequency deviation such as low-orbit satellites. Therefore, it is necessary to study the channel estimation and signal detection schemes for satellite OTFS communication systems. The present invention proposes a DNN-based OTFS time-frequency situational awareness and estimation method. By designing a neural network based on data and model drive, iterative training is performed multiple times to solve the optimal frequency domain correlation coefficient under the minimum mean square error of the network, and the corresponding channel response is calculated, thereby achieving better channel estimation performance.

[0004] According to the literature on OTFS channel estimation found in public channels at home and abroad, the research results of existing literature are mainly in the following aspects: "Pilot sequence design and channel estimation based on orthogonal time-frequency-space (OTFS) system" and other literatures propose a channel estimation algorithm based on the autocorrelation characteristics of the pilot sequence. The performance of this algorithm is related to the PN sequence length. Only a longer PN sequence can guarantee a higher channel estimation accuracy, but a longer sequence length will bring higher computational complexity, and this scheme cannot be applied to communication scenarios where the channel state changes rapidly. References such as "A novel channel estimation scheme for OTFS" propose a threshold-based estimation algorithm, which mainly optimizes the channel estimation performance by setting one or more thresholds. It aims to deal with the channel estimation problem under low signal-to-noise ratio conditions. A Doppler domain transmission pulse pilot is set at the transmitter, and pulse detection is performed on the receiver based on the threshold method. The delay-Doppler domain channel is estimated by the Doppler domain offset, pulse amplitude and number. However, this method has a large pilot overhead and low spectrum utilization. References such as "Research on OTFS Channel Estimation Algorithm in High-Speed ​​Mobile Communication System" propose an estimation method based on compressed sensing, which uses the principle of compressed sensing to estimate the delay, Doppler frequency shift and channel gain of each path, and then determines the size of the pilot transmission matrix to improve the channel estimation accuracy and diversity order. However, this method only considers integer Doppler frequency shift, and does not consider the impact of fractional Doppler frequency shift on the OTFS channel estimation performance.

[0005] With the rapid development of deep learning, related technologies have been applied to communication fields such as modulation recognition and channel equalization. The "OTFS channel estimation based on model-driven deep learning" uses a model-driven deep learning LDAMP algorithm to estimate the OTFS delay Doppler channel, the "OTFS system channel estimation method based on CNN in satellite-to-ground scenarios" uses a convolutional neural network to process the OTFS system in an end-to-end manner, and the "OTFS system signal detection method based on DNN" uses a deep learning method to estimate CSI and directly recover the transmission symbols. However, these methods have poor performance in frequency domain selective fading environments, and directly estimate the received signal, which to a certain extent causes a reduction in transmission efficiency and an increase in complexity. At present, there is no OTFS time-frequency domain channel estimation algorithm that balances high performance and low complexity to achieve high-reliability communication transmission. Therefore, the method proposed in the present invention has obvious novelty characteristics. Summary of the invention

[0006] The present invention proposes an OTFS time-frequency domain situation awareness and channel estimation method based on DNN. The method models the frequency domain selective fading channel in the wireless communication process as an autoregressive process, and then transforms the estimation problem into an autoregressive coefficient estimation process. A neural network driven by data and model is designed, and the neural network is trained by a large amount of prior channel data. The optimal parameter estimation of the network under the minimum mean square error is solved through multiple iterative training, thereby achieving better estimation performance.

[0007] The technical solution of the present invention is: a DNN-based OTFS time-frequency domain situation awareness and channel estimation method, comprising the following steps:

[0008] 1) Based on the channel impulse response function, the pre-processed training data is obtained by performing discrete sampling, DFT transformation, feature matrix representation, and channel prior information generation;

[0009] 2) Design a neural network model and obtain the optimal frequency domain correlation coefficient based on offline training;

[0010] 3) The channel response is estimated using the LS algorithm, and the optimal frequency domain channel response is obtained based on online estimation;

[0011] 4) The received signal sequence is obtained by using the least squares method, and after demodulation and decoding, it is compared with the information sequence at the transmitting end to obtain the bit error rate curve.

[0012] Furthermore, the step 1) acquires training data and pre-processes it according to the OTFS system model;

[0013] a) Taking an OTFS symbol as an example, assuming that the channel response in the symbol remains unchanged, the channel impulse response is converted to s Discretized into g i (n) is expressed as

[0014]

[0015] In the above formula, α l is the average gain of the channel on the lth path; ignoring the inter-symbol interference in the system, n l is the channel delay on the lth path, satisfying n l =[τ l / T s ].

[0016] b) Assume that there are N subcarriers in an OTFS symbol. p After uniformly inserting pilot subcarriers, the discrete impulse response of the system is transformed by N-point DFT, and the channel frequency domain response on the kth subcarrier can be obtained as follows:

[0017]

[0018] c) The channel response vector on each pilot subcarrier within one OTFS symbol time is:

[0019] h=[h(1),…,h(N p )] T

[0020] d) According to the channel response vector, first obtain the eigenvalue D of h N×1 and the corresponding column eigenvector matrix U N×N , arrange the eigenvalues ​​in descending order to get D' N×1 , the corresponding eigenvectors are arranged by index to get U' N×N . Take the first cp eigenvalues ​​and get the representation matrix as follows

[0021] if k≤cp delta(k)=D'(k) / (D'(k)+β / Noise);

[0022] else delta(k)=0;

[0023] e) Generate channel prior information as DNN training data through the channel representation matrix, the expression is as follows:

[0024] Ch_feature=U'*diag(delta)*U' T

[0025] Furthermore, the step 2) obtains channel estimation parameters based on the neural network training model

[0026] A deep neural network (DNN) is a complex artificial neural network structure composed of multiple layers of neurons, each of which is responsible for receiving input, processing, and generating output. The situation awareness and estimation process is mainly divided into two parts: offline training and online prediction.

[0027] In the offline training phase, the CSI at the pilot and the CSI at one OTFS symbol are used as training data, the CSI at the pilot is used as the input data of the DNN, and the CSI at one OTFS symbol is used as the label data to train the DNN;

[0028] The neural network mainly consists of an input layer, a hidden layer and an output layer.

[0029] 1) Take the CFR vector at the pilot obtained by LS estimation as input data, that is,

[0030]

[0031] 2) Preprocess the input data, extract the real and imaginary parts of each CFR vector component and connect them together to get the actual input data

[0032] 3) The nonlinear transformation of the weighted sum of all data in the previous layer is used as the neuron output of the hidden layer. The transformation expression is:

[0033]

[0034] Among them, l l,i , w l and b are the output, weight, and bias of the ith neuron in the lth hidden layer, respectively. T () is the activation function, mainly using the Tanh activation function, the expression is as follows:

[0035]

[0036] 4) The transformation of the neural network can be expressed by the cascade mathematical expression as follows:

[0037]

[0038] Among them, L represents the number of network layers in the neural network, θ represents all the parameters involved in the network, Represents the result of channel estimation.

[0039] 5) Use the Y and channel estimation results output by the OTFS communication system Based on the least squares (LS) channel estimation algorithm, the received sequence r is obtained n :

[0040]

[0041] Furthermore, in step 2), the frequency domain correlation coefficient tmps is obtained by offline training of the DNN neural network. n ;

[0042] The DNN neural network structure is designed as: k hidden layers and 1 output layer. Assuming the input is Ptrain, the neural network training stage can be expressed as follows:

[0043] tmps1=Ptrain*W1+B1

[0044] tmps i =tmps i-1 *W i-1 +BI i-1

[0045] Where i = 2, 3, 4, ... n, the loss error in the training process is obtained as Ptrain-tmps n , used to update W i and B.I. i Equal weights, set the learning rate Lr, the expression is as follows:

[0046] W i =W i +Lr*error

[0047] BI i =BI i +Lr*error

[0048] The number of iterations is set to epoch, and the final output layer expression is as follows:

[0049] WO = WO + Lr * error;

[0050] BO=BO+Lr*error;

[0051] Therefore, the frequency domain correlation coefficient can be calculated as tmps n =tmps n-1 *WO+BO.

[0052] Furthermore, in step 3), the frequency domain channel response H is obtained by using the channel response and the frequency domain correlation coefficient in the online estimation stage. tmp .

[0053] 1) The pilot sequence can be used to obtain the initial estimate H using LS in =r_pilot / Data_pilot, and the channel response H is obtained. in ;

[0054] 2) Using the frequency domain correlation coefficient tmps6 obtained by deep learning, the channel response H is calculated tmp =H in *tmps6.

[0055] 3) Using least squares (LS) estimation to obtain the received signal sequence r n for

[0056] r n =Y*H tmp

[0057] The received signal r n Perform QPSK demodulation and Turbo decoding to finally obtain the received signal sequence R n ,By comparing the information sequences of the receiving end and the sending end, the bit error rate curve and algorithm running time under the situation awareness and estimation algorithm are obtained.

[0058] The advantages of the present invention compared with the prior art are:

[0059] (1) This technology proposes a DNN-based OTFS time-frequency domain situational awareness and channel estimation method. This method obtains training data through the process of discretized sampling, DFT transformation, feature matrix representation, and channel prior information generation based on the channel impulse response function, rather than directly using pilot data or simply normalized data as training data. This method can ensure better channel estimation performance under the premise of lower complexity.

[0060] (2) The DNN-based OTFS time-frequency domain situational awareness and channel estimation method proposed by this technology is mainly divided into offline training and online estimation processes. In the offline training stage, the channel data obtained by pilot sequence processing is used as training data. In the online estimation stage, the channel response is obtained through the trained channel estimation model. Finally, the least squares estimation algorithm is used to obtain the received sequence signal instead of directly estimating the received signal. This technology can achieve lower bit error rate and complexity of communication under the same signal-to-noise ratio conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 It is a flow chart of the present invention.

[0062] Figure 2 It is the neural network channel estimation process.

[0063] Figure 3 The present invention compares the information sequences of the receiving end and the transmitting end of LS estimation, MMSE estimation, LMMSE estimation, and DNN-based time-frequency domain situation awareness and estimation algorithm to obtain bit error rate curves under different channel estimation algorithms. DETAILED DESCRIPTION

[0064] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0065] (1) According to the constructed OTFS system model, the training data is obtained and preprocessed;

[0066] a) Taking an OTFS symbol as an example, assuming that the channel response in the symbol remains unchanged, the channel impulse response is converted to s Discretized into g i (n) is expressed as

[0067]

[0068] In the above formula, α l is the average gain of the channel on the lth path; ignoring the inter-symbol interference in the system, n l is the channel delay on the lth path, satisfying n l =[τ l / T s ].

[0069] b) Assume that the number of subcarriers N in an OTFS symbol is set to 64 and the number of pilot subcarriers N p Set to 4, after uniformly inserting the pilot, perform N-point DFT transformation on the discrete impulse response of the system, and the channel frequency domain response on the kth subcarrier can be obtained as

[0070]

[0071] c) The channel response vector on each pilot subcarrier within one OTFS symbol time is:

[0072] h=[h(1),…,h(N p )] T

[0073] d) According to the channel response vector, first obtain the eigenvalue D of h N×1 and the corresponding column eigenvector matrix U N×N , arrange the eigenvalues ​​in descending order to get D' N×1 , the corresponding eigenvectors are arranged by index to get U' N×N . Take the first cp eigenvalues ​​and get the representation matrix as follows

[0074] if k≤cp delta(k)=D'(k) / (D'(k)+β / Noise);

[0075] else delta(k)=0;

[0076] e) Generate channel prior information as DNN training data through the channel representation matrix, the expression is as follows:

[0077] Ch_feature=U'*diag(delta)*U' T

[0078] (2) Use DNN neural network training to obtain the frequency domain correlation coefficient tmps6;

[0079] The DNN neural network structure is designed as follows: 4 hidden layers and 1 output layer. Assuming the input training data Ptrain2, the neural network training stage can be expressed as follows:

[0080] tmps1=Ptrain2*W1+B1

[0081] tmps i =tmps i-1 *W i-1 +BI i-1

[0082] Where i = 2, 3, 4, 5, 6, the loss error in the training process is Ptrain2-tmps6, which is used to update W i and B.I. i With equal weights, set the learning rate Lr = 0.005, and the expression is as follows:

[0083] W i =W i +Lr*error

[0084] BI i =BI i +Lr*error

[0085] The number of iterations is set to epoch = 2000, and the final output layer expression is as follows:

[0086] WO = WO + Lr * error;

[0087] BO=BO+Lr*error;

[0088] Therefore, the frequency domain correlation coefficient can be calculated as tmps6=tmps5*WO+BO.

[0089] In the step 3) in the online estimation stage, the frequency domain channel response H is obtained by using the channel response and the frequency domain correlation coefficient. tmp .

[0090] 1) The pilot sequence can be used to obtain the initial estimate H using LS in =r_pilot / Data_pilot, and the channel response H is obtained. in ;

[0091] 2) Using the frequency domain correlation coefficient tmps6 obtained by deep learning, the channel response H is calculated tmp =H in *tmps6.

[0092] 3) Using least squares (LS) estimation to obtain the received signal sequence r n for

[0093] r n =Y*H tmp

[0094] The present invention mainly adopts simulation data for verification, and all steps and conclusions are verified to be correct on Matlab.

[0095] By comparing the information sequences of the receiver and transmitter of LS estimation, MMSE estimation, LMMSE estimation, and DNN-based time-frequency domain situation awareness and estimation algorithm, the bit error rate curves under different channel estimation algorithms are obtained. Figure 3 .

[0096] Table 1 Comparison of running time of different algorithms

[0097] algorithm LS channel estimation MMSE channel estimation IMMSE channel estimation DNN-based channel estimation Running time / s 176.796548 1172.333143 968.738319 930.13672

[0098] The bit error rate can reflect the channel estimation performance. The lower the bit error rate, the more accurate the estimated received sequence is and the better the channel estimation algorithm performance is. The algorithm running time reflects the algorithm complexity to a certain extent. The shorter the algorithm running time, the lower the algorithm complexity is.

[0099] Figure 3 The performance of LS estimation, MMSE estimation, IMMSE estimation, and DNN-based time-frequency domain situation awareness and estimation algorithm is compared with Table 1. It can be seen that although the LS estimation algorithm has a lower computational complexity, its estimation performance is poor, resulting in poor anti-interference capability; the MMSE estimation algorithm and IMMSE estimation algorithm have better performance, but at the same time they need to obtain channel prior information and have higher computational complexity; the performance of the DNN-based time-frequency domain situation awareness and estimation algorithm is between the MMSE algorithm and the IMMSE algorithm, but the computational complexity is greatly reduced, and when SNR≥6dB, the bit error rate can reach 10 -4 , it has a good balance between algorithm complexity and estimation performance and has good anti-interference ability.

[0100] The contents not described in detail in the specification of the present invention belong to the common knowledge of those skilled in the art.

Claims

1. A DNN-based OTFS time-frequency domain situation awareness and channel estimation method, characterized in that: include: 1) Based on the channel impulse response function, discretization sampling, DFT transformation, feature matrix representation, and channel prior information generation process are performed to obtain pre-processed training data; 2) Design a neural network model and obtain the optimal frequency domain correlation coefficient based on offline training; 3) The channel response is estimated using the LS algorithm, and the optimal frequency domain channel response is obtained based on online estimation; 4) The received signal sequence is obtained by using the least squares method, and after demodulation and decoding, it is compared with the information sequence at the transmitting end to obtain the bit error rate curve.

2. According to the DNN-based OTFS time-frequency domain situation awareness and channel estimation method according to claim 1, it is characterized in that: The step 1) acquires training data and performs preprocessing according to the OTFS system model. The specific process is as follows: a) The channel impulse response is sampled at a sampling interval T s Discretized into g i (n) is expressed as: α l is the average gain of the channel on the lth path; ignoring the inter-symbol interference in the system, n l is the channel delay on the lth path, satisfying n l =[τ l / T s ]; b) Assume that there are N subcarriers in an OTFS symbol. p After uniformly inserting pilot subcarriers, the discrete impulse response of the system is transformed by N-point DFT, and the channel frequency domain response on the kth subcarrier is obtained as follows: c) The channel response vector on each pilot subcarrier within one OTFS symbol time is obtained as: h=[h(1),…,h(Np)] T d) According to the channel response vector, first obtain the eigenvalue D of h N×1 and the corresponding column eigenvector matrix U N×N , arrange the eigenvalues ​​in descending order to get D' N×1 , the corresponding eigenvectors are arranged by index to get U' N×N ; Take the first cp eigenvalues ​​and obtain the representation matrix as follows: if k≤cp delta(k)=D'(k) / (D'(k)+β / Noise); else delta(k)=0; e) Generate channel prior information as DNN training data through the channel representation matrix, the expression is as follows: Ch_feature=U'*diag(delta)*U' T 。 3. The DNN-based OTFS time-frequency domain situation awareness and channel estimation method according to claim 1, characterized in that: The step 2) uses DNN to design a neural network for offline training to obtain the frequency domain correlation coefficient tmps n ; The DNN neural network structure is designed as: k hidden layers and 1 output layer. Assuming the input is Ptrain, the neural network training stage can be expressed as follows: tmps1=Ptrain*W1+B1 <h2 style=";text-align:left;direction:ltr">tmps<h2 style=";text-align:left;direction:ltr"> i <h2 style=";text-align:left;direction:ltr"> =tmps<h2 style=";text-align:left;direction:ltr"> i-1 <h2 style=";text-align:left;direction:ltr"> *W<h2 style=";text-align:left;direction:ltr"> i-1 <h2 style=";text-align:left;direction:ltr"> +BI<h2 style=";text-align:left;direction:ltr"> i-1 Where i = 2, 3, 4, ... n, the loss error in the training process is obtained as Ptrain-tmps n , used to update W i and B.I. i Equal weights, set the learning rate Lr, the expression is as follows: W i =W i +Lr*error BI i =BI i +Lr*error The number of iterations is set to epoch, and the final output layer expression is as follows: WO=WO+Lr*error; BO=BO+Lr*error; Therefore, the frequency domain correlation coefficient can be calculated as tmps n =tmps n-1 *WO+BO.

4. The DNN-based OTFS time-frequency domain situation awareness and channel estimation method according to claim 1, characterized in that: In the step 3) in the online estimation stage, the frequency domain channel response H is obtained by using the channel response and the frequency domain correlation coefficient. tmp , the calculation method is as follows: 1) The pilot sequence can be used to obtain the initial estimate H using LS in =r_pilot / Data_pilot, and the channel response H is obtained. in ; 2) Using the frequency domain correlation coefficient tmps6 obtained by deep learning, the channel response H is calculated tmp =H in *tmps6; 3) Using least squares (LS) estimation to obtain the received signal sequence r n for: r n =Y*H tmp 。

Images