CN116232810A - A OTFS Channel Estimation Method Based on Deep Neural Network - Google Patents

Abstract

The invention discloses an OTFS channel estimation method based on a deep neural network. The method includes: based on the zero-padding OTFS structure, using the traditional pilot scheme with protection symbols, constructing an OTFS transceiver system, generating a data set of receiving pilot signals and corresponding channel information; building a 4-layer depth that meets the input and output sizes Neural network; the pilot received in the OTFS domain is used as the input of the network, and the corresponding channel parameters are used as the label training network, and the parameters of the deep neural network are adjusted to construct the DNN model with the best performance; the transmission data is generated online, and the trained neural network is used The network estimates the channel; using the MRC detector and the estimated channel parameters, signal detection is performed on the transmitted data. Compared with the traditional channel estimation method, the present invention improves the accuracy of channel estimation with lower pilot power, obtains the bit error rate performance close to the ideal channel, and the accuracy and generalization ability of the DNN network guarantees the high-speed mobile environment The accuracy of channel estimation in medium.

Description

OTFS channel estimation method based on deep neural network

Technical Field

The invention relates to the technical field of wireless communication, in particular to an OTFS channel estimation method based on a deep neural network.

Background

In recent years, rapid development of high-speed trains, low-Orbit Satellites (LEOs), and the like has been greatly focused on how to achieve high-quality reliable communication in these high-speed motion scenes. Because of the time-varying channel and high doppler shift in these scenarios, the difficulty of estimating the time-frequency dual-choice channel is greatly increased, and the orthogonal time-frequency space (OTFS) is generated as a two-dimensional modulation in the delay-doppler domain, which converts the signal transmission in the conventional time-frequency domain into the time-independent delay-doppler domain, making full use of the time and frequency diversity of the channel. From the realization point of view, the OTFS modulation only needs to add a preprocessing module and a post-processing module on the basis of an OFDM modulation frame, and has great engineering value and application potential.

Since the reflectors in the propagation path are limited, the number of channel paths is limited. Meanwhile, the effective channel in the delay-doppler domain is sparse, and many efficient signal detection schemes, such as LMMSE, message passing algorithm, MRC detection, and variational bayesian method, are studied according to this channel characteristic. However, to ensure reliable signal detection, more accurate channel state information is also required, and current research is mainly focused on integer doppler channels, such as orthogonal matching pursuit technology, embedded single pilot delay-time domain channel estimation, message-based fractional doppler channel estimation, and the like. In OTFS, it is generally assumed that the data within a frame is subject to the same channel, so it is required that the duration of a frame not be too long, i.e. N not too large, which results in doppler resolution

The fractional doppler effect is more pronounced than small enough, especially in some high speed scenarios, such as LEO communications and Unmanned Aerial Vehicle (UAV) communications.

The machine learning technology has wide research and practical application in the fields of image processing, voice recognition and the like, is increasingly combined with wireless communication, and has more obvious advantages compared with the traditional method in the aspects of channel estimation, signal detection and the like. The function of fitting the corresponding target value by giving out a group of similar signals is realized by extracting signal characteristics and continuously training and learning, and in addition, the strong learning capacity and flexible network structure of the method have important effects on improving the estimation precision.

Disclosure of Invention

The invention aims to provide an OTFS channel estimation method based on a deep neural network, which can reduce pilot frequency power and improve OTFS fractional Doppler channel estimation performance.

The technical solution for realizing the purpose of the invention is as follows: an OTFS channel estimation method based on a deep neural network comprises the following steps:

step 1, based on an OTFS structure of zero padding, constructing an OTFS receiving and transmitting system by using a traditional pilot frequency scheme with a protection symbol and generating a data set for receiving pilot frequency signals and corresponding channel information;

step 2, building a 4-layer deep neural network conforming to the input and output sizes;

step 3, pilot frequency received in the OTFS domain is used as input of a network, corresponding channel parameters are used as a label training network, and parameters of the deep neural network are adjusted to construct a DNN model with optimal performance;

step 4, generating transmission data on line, and estimating a channel by using the trained neural network;

and 5, performing signal detection on the transmission data by using the MRC detector and the estimated channel parameters.

Further, the OTFS structure based on zero padding described in step 1 uses a conventional pilot scheme with guard symbols to construct an OTFS transceiver system and generate a data set of a received pilot signal and corresponding channel information, which is specifically as follows:

step 1.1, bit information flows are mapped into corresponding transmitting information through modulation, the transmitting information is placed on a time delay-Doppler domain grid according to a certain rule, the transmitting information comprises data symbols and pilot symbols, the pilot symbols are placed in a stack of zero padding symbols, the zero padding symbols play the same role as the traditional protection symbols, and the transmitting information is specifically as follows:

wherein ,L_ZP For the length of the line where the zero padding symbol is located, in order to ensure that the data and the pilot frequency are not interfered with each other, the neural network is beneficial to extracting the characteristics of the received pilot frequency, and L is taken _ZP ＝2l _m +1，l _m For maximum value of delay tap, the pilot position is generally set as

Step 1.2 the OTFS transceiver system comprises mapping a bit information stream onto constellation symbols to form a delay-Doppler signal, an inverse fast Fourier transform (ISFFT) module, a Hemson Barbell transform module, a time-varying channel, a Wigner transform module, and an even fast Fourier transform (SFFT) module, according to the transmission data structure of step 1.1, when it is subjected to a signal with a fixed number of subcarriers M, a number of symbols N, a carrier frequency f _c And subcarrier spacing delta f, and the receiving end will receive the delay-Doppler pilot information Y _p Collected as a data set, corresponding channel information P _ch ＝{h _i ,k _i ,l _i P, 1.ltoreq.i.ltoreq.P as a tag, where h _i 、k _i 、l _i The channel gain, doppler tap index and delay tap index of the ith path, respectively, P is the number of propagation paths, and in order to obtain a more accurate network we use as many channel conditions as possible to generate data sets, such as different delays, channel gains, signal to noise ratios and doppler shifts.

Further, the step 2 of building a 4-layer deep neural network conforming to the input and output sizes is specifically as follows:

the deep neural network of the method has 4 layers, the first layer is an input layer, the input data is the real part and the imaginary part of the time delay-Doppler domain receiving pilot signal, and therefore the number of neurons is 2 (l) _m +1) N, fourth layer is lineAn sexual output layer for outputting data as estimated channel parameters

Since 1.ltoreq.i.ltoreq.P, the output information is split into a real part and an imaginary part for output, respectively, considering that the channel gain is complex, and thus the number of neurons is 2 (1+3P), and P is the number of propagation paths. The middle two layers are hidden layers, and the number of neurons is Q respectively ₁ 、Q ₂ The hidden layers are all full-connection layers, and the activation function selects the Relu function.

Further, in step 3, the pilot frequency received in the OTFS domain is used as an input of the network, the corresponding channel parameter is used as a label training network, and parameters of the deep neural network are adjusted to construct a DNN model with optimal performance, which is specifically as follows:

step 3.1, dividing the data set into a training set T and a verification set V according to the data set in the step 1, performing digital orthogonal transformation, obtaining a real part and an imaginary part of a signal, and splicing the real part and the imaginary part into an input sample of a network;

and 3.2, setting key network parameters including Batch size, learning rate, training round number epoch, and the like according to the deep neural network in the step 2, performing network training by using a training set T, supervising a verification set V, enabling gradient to be reduced by adopting a self-adaptive moment estimation (Adam) optimization algorithm to continuously adjust network weight and bias, optimizing the network, judging whether the network meets the requirement or not by calculating a Mean Square Error (MSE) loss function, adjusting the network parameters to perform training if the loss function value does not meet the requirement (the loss function value is not small enough), obtaining a network initial model, then testing network estimation precision on line, and performing training again by modifying the neural network model, adjusting the network parameters to obtain a neural network model with optimal performance and storing the neural network model.

Further, the online generation of the transmission data in step 4 uses the trained neural network to estimate the channel, which is specifically as follows:

step 4.1, mapping the randomly generated bit information stream into constellation diagram symbols according to the step 1, and discharging the constellation diagram symbols to a delay-Doppler grid according to a certain rule to generate transmission data X _DD Through OTFS conversionTo a time domain transmitting signal s (t) and transmitting to a time-varying channel, a time domain signal r (t) received by a receiving end is:

wherein ω (t) is obeyed to have a mean value of 0 and a variance of σ ² H (τ, v) is the channel impulse response of the delay-doppler domain;

step 4.2, the time domain received signal r (t) is transmitted by two paths, one path is transformed to the delay-Doppler domain to obtain Y _DD For channel estimation, the other path is used for sampling interval

Sampling and dispersing the signals which are more than or equal to 0 and less than or equal to MN-1, and then taking the samples and the discretized time domain signals as the input of a signal detector, wherein the discrete time domain signals are as follows:

wherein ,

is a time domain channel matrix []Is a mould-taking operation;

step 4.3, according to the network model saved in the step 3, Y is calculated _DD Is input to the neural network by recombination of the real part and the imaginary part of the channel information to obtain estimated channel information

And finishing the fractional Doppler channel estimation.

Further, the signal detection of the transmission data using the MRC detector and the estimated channel parameters in step 5 is specifically as follows:

step 5.1, reconstructing a channel matrix g according to the method of the step 4;

step 5.2, according to the method of step 4, signal detection is performed by using a time delay-time domain MRC detection algorithm, and the output vector of the MRC is:

wherein .

Another representation of the transmitted data, the residual interference of the maximum combining ratio and the time domain channel g, respectively, of the time delay-time domain +.>

Representing element division and element multiplication, and finally solving the estimated signal according to the maximum likelihood criterion

The method comprises the following steps:

wherein A is constellation diagram, F _N Is a normalized N-point discrete fourier transform.

Compared with the prior art, the invention has the remarkable advantages that: since OTFS channels are parameterized by channel gain and fractional doppler shift, these parameters are estimated directly by learning the characteristics of the received pilot, rather than estimating the channel matrix directly, reducing estimation errors by estimating fewer variables; the diversity training network of the data set is increased, stronger robustness and generalization capability are realized, and the performance under various high-mobility scenes is ensured; simulation shows that the method realizes more accurate channel estimation under lower pilot frequency expense, and improves the detection performance of the OTFS system under the fractional Doppler channel.

Drawings

Fig. 1 is a flow chart of an OTFS channel estimation method based on a deep neural network.

Fig. 2 is a schematic diagram of transmission data based on the zero padding OTFS structure.

Fig. 3 is a schematic structural diagram of a neural network.

Fig. 4 is a NMSE performance diagram under two propagation paths.

Fig. 5 is a BER performance diagram under two propagation paths.

Fig. 6 is a graph of BER performance for different doppler shifts under two propagation paths.

Fig. 7 is a graph of BER performance under single path conditions.

Detailed Description

The invention discloses an OTFS channel estimation method based on a deep neural network, which comprises the following steps:

step 1, based on an OTFS structure of zero padding, constructing an OTFS receiving and transmitting system by using a traditional pilot frequency scheme with a protection symbol and generating a data set for receiving pilot frequency signals and corresponding channel information;

step 2, building a 4-layer deep neural network conforming to the input and output sizes;

step 3, pilot frequency received in the OTFS domain is used as input of a network, corresponding channel parameters are used as a label training network, and parameters of the deep neural network are adjusted to construct a DNN model with optimal performance;

step 4, generating transmission data on line, and estimating a channel by using the trained neural network;

and 5, performing signal detection on the transmission data by using the MRC detector and the estimated channel parameters.

Further, the OTFS structure based on zero padding described in step 1 uses a conventional pilot scheme with guard symbols to construct an OTFS transceiver system and generate a data set of a received pilot signal and corresponding channel information, which is specifically as follows:

step 1.1, bit information flow is mapped into corresponding transmitting information through modulation, and is placed on a grid of a time delay-Doppler domain according to a certain rule

k∈[0,N-1],l∈[0,M-1]In general, tΔf=1, and the transmission information includes data symbols and pilot symbols, as shown in fig. 2, the pilot symbols are placed in a stack of zero padding symbols, and the zero padding symbols play the same role as the conventional guard symbols, and the transmission information is specifically as follows:

wherein ,L_ZP For the length of the line where the zero padding symbol is located, in order to ensure that the data and the pilot frequency are not interfered with each other, the neural network is beneficial to extracting the characteristics of the received pilot frequency, and L is taken _ZP ＝2l _m +1，l _m For maximum value of delay tap, the pilot position is generally set as

Step 1.2 the OTFS transceiver system as shown in FIG. 1, comprises mapping bit information stream onto constellation symbols to form delay-Doppler signal, inverse even fast Fourier transform (ISFFT) module, hessenberg transform module, time-varying channel, wigner transform module and even fast Fourier transform (SFFT) module, according to the transmission data structure of step 1.1, when it has a fixed subcarrier number M, symbol number N, carrier frequency f _c And subcarrier spacing delta f, and the receiving end will receive the delay-Doppler pilot information Y _p Collected as a data set, corresponding channel information P _ch ＝{h _i ,k _i ,l _i P, 1.ltoreq.i.ltoreq.P as a tag, where h _i 、k _i 、l _i The channel gain, doppler tap index and delay tap index of the ith path, respectively, P is the number of propagation paths, and in order to obtain a more accurate network we use as many channel conditions as possible to generate data sets, such as different delays, channel gains, signal to noise ratios and doppler shifts.

Further, the step 2 of building a 4-layer deep neural network conforming to the input and output sizes is specifically as follows:

the deep neural network of the present invention has 4 layers in total, as shown in FIG. 3, the first layer is the input layer, the input data is the real part and the imaginary part of the delay-Doppler domain received pilot signal, and thus the number of neurons is 2 (l) _m +1) N, the fourth layer is a linear output layer, and the output data is estimated channel parameters

Since the channel gain is complex, the output information is split into a real part and an imaginary part and output, respectively, and thus the number of neurons is 2 (1+3p), and P is the number of propagation paths. The middle two layers are hidden layers, and the number of neurons is Q respectively ₁ 、Q ₂ The hidden layers are all full-connection layers, and the activation function selects the Relu function.

Further, in step 3, the pilot frequency received in the OTFS domain is used as an input of the network, the corresponding channel parameter is used as a label training network, and parameters of the deep neural network are adjusted to construct a DNN model with optimal performance, which is specifically as follows:

step 3.1, dividing the data set into a training set T and a verification set V according to the data set in the step 1, performing digital orthogonal transformation, obtaining a real part and an imaginary part of a signal, and splicing the real part and the imaginary part into an input sample of a network;

and 3.2, setting key network parameters including Batch size, learning rate, training round number epoch, and the like according to the deep neural network in the step 2, performing network training by using a training set T, supervising a verification set V, enabling gradient to be reduced by adopting a self-adaptive moment estimation (Adam) optimization algorithm to continuously adjust network weight and bias, optimizing the network, judging whether the network meets the requirement or not by calculating a Mean Square Error (MSE) loss function, adjusting the network parameters to perform training if the loss function value does not meet the requirement (the loss function value is not small enough), obtaining a network initial model, then testing network estimation precision on line, and performing training again by modifying the neural network model, adjusting the network parameters to obtain a neural network model with optimal performance and storing the neural network model.

Further, the online generation of the transmission data in step 4 uses the trained neural network to estimate the channel, which is specifically as follows:

step 4.1, mapping the randomly generated bit information stream into constellation diagram symbols according to the step 1, and discharging the constellation diagram symbols to a delay-Doppler grid according to a certain rule to generate transmission data X _DD The OTFS conversion is carried out to obtain a time domain transmitting signal s (t) and the time domain transmitting signal s (t) is transmitted to a time-varying channel, and the time domain signal r (t) received by a receiving end is:

r(t)＝∫∫h(τ,ν)e ^j2πν(t-τ) s(t-τ)dτdν+ω(t) (2)

wherein ω (t) is obeyed to have a mean value of 0 and a variance of σ ² H (τ, v) is the channel impulse response of the delay-doppler domain;

step 4.2, the time domain received signal r (t) is transmitted by two paths, one path is transformed to the delay-Doppler domain to obtain Y _DD For channel estimation, the other path is used for sampling interval

Sampling and dispersing the signals which are more than or equal to 0 and less than or equal to MN-1, and then taking the samples and the discretized time domain signals as the input of a signal detector, wherein the discrete time domain signals are as follows:

wherein ,

is a time domain channel matrix, []Is a mould-taking operation;

step 4.3, according to the network model saved in the step 3, Y is calculated _DD Is input to the neural network by recombination of the real part and the imaginary part of the channel information to obtain estimated channel information

And finishing the fractional Doppler channel estimation.

Further, the signal detection of the transmission data using the MRC detector and the estimated channel parameters in step 5 is specifically as follows:

step 5.1, reconstructing a channel matrix g according to the method of step 4, and expressing the time-frequency domain index as q=m+nm, where the channel matrix may be expressed as:

step 5.2, according to the method of step 4, signal detection is performed by using a time delay-time domain MRC detection algorithm, and the output vector of the MRC is:

wherein .

Another representation of the transmitted data, the residual interference of the maximum combining ratio and the time domain channel g, respectively, of the time delay-time domain +.>

Representing the element division and the element multiplication, and finally solving the estimated signal according to the maximum likelihood criterion>

The method comprises the following steps:

wherein A is constellation diagram, F _N Is a normalized N-point discrete fourier transform.

The invention will be described in further detail with reference to the drawings and the specific examples.

Examples

The use scene considered by the invention is satellite communication and unmanned aerial vehicle communication, the channel refers to a model in NTN-TDL series in 3GPP standard, the maximum speed of a mobile terminal is set to 1000km/h, the maximum relative time delay is about 4us, and the corresponding time delay tap is l _m =4. Because of the small number of paths in UAV communications, the channel is modeled as a two-path model. The OTFS system parameters were set as follows: m=64, n=8, f _c =3 GHz, Δf=15 kHz, and 4-QAM modulation scheme is used. For ease of analysis to illustrate the performance of the present invention, the power ratio of pilot and data symbols is defined as:

since there are some guard symbols near the pilot, the energy of the pilot can be increased by borrowing the energy of the guard symbols. In addition, the normalized mean square error is defined as:

according to the method described in step 1, 150000 frames of data are generated as data sets, wherein the training set T has 147000 frames and the verification set has 3000 frames. According to the deep neural network in the step 2, batch size Batch size=64, learning rate Learning rate=0.001, and training round number epochs=20 are set. According to the training network mode in the step 3, the training set T is used for carrying out network training, the verification set V is used for supervision to obtain a network initial model, then the network estimation precision is tested on line, and the training is carried out again by modifying the neural network model and adjusting network parameters to obtain the neural network model with optimal performance. Wherein the performance of both concealment layers is better than any other number of concealment layers, because too many layers increase the computational complexity and the performance does not increase accordingly, while too few layers do not estimate the channel accurately. In addition, when Q ₁ ＝256、Q ₂ At=128, the performance is synthetically optimal. Finally, the best performance is achieved when the training signal to noise ratio is 10dB and the activation function of the hidden layer is Relu. And (3) according to the online channel estimation mode in the step (4), 3000 frames of data are sent for testing.

Fig. 4 compares NMSE performance under two propagation paths for a Deep Neural Network (DNN) based estimation method with a conventional single pilot method. As can be seen from the figure, the performance of the NMSE will increase with increasing pilot power. The pilot power of the traditional method is far higher than the data power, so that the performance similar to that of the method proposed by us can be achieved. For example, a DNN-based method using one pilot symbol is shown at E _p /E _s The traditional method can be realized at E when the frequency is 14dB _p /E _s 30dB performance. The method is characterized in that the traditional estimation method based on the single pilot frequency directly estimates partial channels under the common influence of channel gain and Doppler frequency shift, and finally obtains the whole time domain channel matrix by utilizing an interpolation method, so that the estimation error is larger. And directly estimated using DNN-based estimation methodsThe related parameters of the channel are calculated, the estimated variables are fewer, and the error of the reconstructed channel matrix is correspondingly smaller. Furthermore, DNN-based methods are less sensitive to noise and contribute to improving estimation accuracy. This has an important role in saving resources and reducing peak-to-average ratio (PAPR). At the same pilot power, we compared the performance of the NMSE of the present invention using one pilot with two pilots, and the results indicate that one pilot has the same estimation accuracy as the two pilots.

Fig. 5 analyzes the performance of the present invention by simulating BER under two propagation paths. The performance of the conventional method is shown by the curve marked with circles, at E _p /E _s When the performance is improved, the performance is obviously improved. While the invention follows E _p /E _s The performance improvement amplitude is less obvious than the conventional method because the DNN-based estimation method curves more closely approximate the bit error rate performance of the known CSI, the conventional method is at E _p /E _s Performance at=30db with the proposed method at E _p /E _s As is the case when=14 dB. In order to make the error rate performance consistent with the ideal situation, the pilot power is required to be 23dB higher than the data power; when conventional method E _p /E _s When the power is increased to 40dB, and E _p /E _s Compared with 30dB, the BER is 10 ^-4 Only about 0.3dB of gain is achieved. Therefore, proper pilot power not only ensures accurate channel estimation, but also reasonably allocates resources. Furthermore, the NMSE and BER performances of both methods have the same conclusion, reflecting the advantages of the present invention.

Fig. 6 shows the generalization of DNN networks under different doppler conditions. It can be observed that the error rates under different doppler conditions are similar except for slight differences, which demonstrates the effectiveness of the DNN-based method at different movement speeds. Finally, we simulate the channel estimation of DNN-based methods under one propagation path, with a single path being a special case where the channel gain of the second path in the two-path model is 0. As shown in fig. 7, the DNN-based method can still save about 16dB of pilot power to achieve the conventional method at E _p /E _s Bit error rate performance under the condition of=30db. Based onDNN method at E _p /E _s The performance of the ideal channel is closer to that of the 23dB, and furthermore, the NMSE in the case of a single path and the BER of different doppler, both demonstrate the effectiveness of the DNN-based approach.

Claims (6)

1. An OTFS channel estimation method based on a deep neural network is characterized by comprising the following steps:

step 1, based on an OTFS structure of zero padding, constructing an OTFS receiving and transmitting system by using a traditional pilot frequency scheme with a protection symbol and generating a data set for receiving pilot frequency signals and corresponding channel information;

step 2, building a 4-layer deep neural network conforming to the input and output sizes;

step 3, pilot frequency received in the OTFS domain is used as input of a network, corresponding channel parameters are used as a label training network, and parameters of the deep neural network are adjusted to construct a DNN model with optimal performance;

step 4, generating transmission data on line, and estimating a channel by using the trained neural network;

and 5, performing signal detection on the transmission data by using the MRC detector and the estimated channel parameters.

2. The OTFS channel estimation method based on deep neural network according to claim 1, wherein the OTFS structure based on zero padding in step 1 uses a conventional pilot scheme with guard symbols to construct an OTFS transceiver system and generate a data set of a received pilot signal and corresponding channel information, specifically as follows:

step 1.1, the bit information stream is mapped into corresponding transmitting information through modulation, wherein the transmitting information comprises data symbols and pilot symbols, the pilot symbols are placed in a pile of zero padding symbols, the zero padding symbols play the same role as the traditional protection symbols, and the transmitting information is specifically as follows:

wherein ,L_ZP for the length of the line where the zero padding symbol is located, in order to ensure that the data and the pilot frequency are not interfered with each other, the neural network is beneficial to extracting the characteristics of the received pilot frequency, and L is taken _ZP ＝2l _m +1，l _m For maximum value of delay tap, the pilot position is generally set as

Step 1.2, the OTFS transceiver system includes mapping the bit information stream onto constellation symbols to form a delay-doppler signal, an inverse fast fourier transform (ISFFT) module, a hessburg transform module, a time-varying channel, a wigner transform module, and an even fast fourier transform (SFFT) module, the transmission data in step 1.1 passes through the OTFS transceiver system, and the receiving end transmits the delay-doppler pilot information Y to the receiving end _p Corresponding channel information

The channel information, including channel gain, doppler tap index, delay tap index and propagation path number, is collected as a data set, which is generated using as many channel conditions as possible, such as different delays, channel gains, signal to noise ratios and doppler shifts, in order to obtain a more accurate network.

3. The OTFS channel estimation method based on a deep neural network according to claim 2, wherein the constructing a 4-layer deep neural network according to the input/output size in step 2 specifically includes:

the deep neural network of the invention has 4 layers, the first layer is an input layer, and the input data is a time delay-Doppler domain received pilot signal Y _p The real and imaginary parts of (2) thus the number of neurons is 2 (l _m +1) N; the fourth layer is a linear output layer, and the output data is estimated channel parameters

Considering that the channel gain is complex, the output information is split into a real part and an imaginary part to be respectively output, so that the number of neurons is 2 (1+3P), and P is the number of propagation paths; in (a)The two layers are hidden layers, and the number of neurons is Q ₁ 、Q ₂ The hidden layers are all full-connection layers, and the activation function selects the Relu function.

4. The OTFS channel estimation method based on the deep neural network according to claims 2 and 3, wherein in step 3, the pilot frequency received in the OTFS domain is used as an input of the network, the corresponding channel parameter is used as a label training network, and parameters of the deep neural network are adjusted to construct a DNN model with optimal performance, which specifically comprises the following steps:

step 3.1, dividing the data set into a training set T and a verification set V according to the data set in the step 1, performing digital orthogonal transformation, obtaining a real part and an imaginary part of a signal, and splicing the real part and the imaginary part into an input sample of a network;

and 3.2, setting key network parameters including Batch size, learning rate, training round number epoch, and the like according to the deep neural network in the step 2, performing network training by using a training set T, supervising a verification set V, enabling gradient to be reduced by adopting a self-adaptive moment estimation (Adam) optimization algorithm to continuously adjust network weight and bias, optimizing the network, judging whether the network meets the requirement or not by calculating a Mean Square Error (MSE) loss function, adjusting the network parameters to perform training if the loss function value does not meet the requirement (the loss function value is not small enough), obtaining a network initial model, then testing network estimation precision on line, and performing training again by modifying the neural network model, adjusting the network parameters to obtain a neural network model with optimal performance and storing the neural network model.

5. The OTFS channel estimation method based on deep neural network according to claims 1 and 4, wherein the online generating of the transmission data in step 4 estimates the channel using a trained neural network, specifically as follows:

step 4.1, mapping the randomly generated bit information stream into constellation diagram symbols according to the step 1, and discharging the constellation diagram symbols to a delay-Doppler grid according to a certain rule to generate transmission data X _DD Obtaining a time domain transmitting signal s (t) through OTFS conversion and transmitting the time domain transmitting signal s (t) to a time-varying channel for receivingThe time domain signal r (t) received by the terminal is:

wherein ω (t) is obeyed to have a mean value of 0 and a variance of σ ² H (τ, v) is the channel impulse response of the delay-doppler domain;

step 4.2, the time domain received signal r (t) is transmitted in two paths, one path being transformed to the delay-Doppler domain, Y _DD For channel estimation, the other path is used for sampling interval

Sampling and dispersing the signals which are more than or equal to 0 and less than or equal to MN-1, and then taking the samples and the discretized time domain signals as the input of a signal detector, wherein the discrete time domain signals are as follows:

wherein ,

is a time domain channel matrix, []Is a mould-taking operation;

step 4.3, according to the network model saved in the step 3, Y is calculated _DD Is input to the neural network by recombination of the real part and the imaginary part of the channel information to obtain estimated channel information

Channel estimation is completed.

6. The OTFS channel estimation method based on deep neural network according to claim 5, wherein the signal detection of the transmission data using the MRC detector and the estimated channel parameters in step 5 is specifically as follows:

step 5.1, reconstructing a channel matrix g according to the method of the step 4;

step 5.2, according to the method of step 4, signal detection is performed by using a time delay-time domain MRC detection algorithm, and the output vector of the MRC is:

wherein ,

Δg _m 、

another representation of the transmitted data, the residual interference of the maximum combining ratio and the time domain channel g, respectively, of the time delay-time domain +.>

Representing element division and element multiplication, and finally solving the estimated signal according to the maximum likelihood criterion

The method comprises the following steps: />

Wherein A is constellation diagram, F _N Is a normalized N-point discrete fourier transform.

Images