CN115987734A

CN115987734A - Low-complexity OTFS system symbol detection method based on deep neural network

Info

Publication number: CN115987734A
Application number: CN202211542551.1A
Authority: CN
Inventors: 刘伟; 于同阳; 白宝明
Original assignee: Xidian University
Current assignee: Xidian University
Priority date: 2022-12-02
Filing date: 2022-12-02
Publication date: 2023-04-18

Abstract

The invention discloses a deep neural network-based OTFS (optical transmission system) system symbol detection method, which mainly solves the problems of high symbol detection complexity and low symbol receiving detection speed in the prior art. The method comprises the following implementation steps: 1. obtaining a training set; 2. constructing a neural network; 3. training the deep neural network by using a training set; 4. receiving a time domain signal sent by a transmitting terminal, and carrying out Virger transformation on the time domain signal to obtain a time-frequency domain signal; 5. carrying out Fourier transform on the time-frequency domain signal to obtain a receiving symbol of a delay-Doppler domain; 6. and detecting the received symbols by adopting the trained deep neural network to obtain the estimated value of the transmitted symbols. The invention reduces the complexity of the whole symbol detection of the OTFS system, improves the speed of detecting the received symbol, and can be used for recovering the transmitted signal from the received signal of the OTFS system.

Description

Low-complexity OTFS system symbol detection method based on deep neural network

Technical Field

The invention belongs to the technical field of communication, and further relates to a symbol detection method of a low-complexity OTFS (optical transport plane) system, which can be used for recovering a transmitting signal from a receiving signal of the OTFS system.

Background

Currently, orthogonal frequency division multiplexing OFDM modulation techniques widely used in 4G, 5G, and WIFI wireless networks are susceptible to the doppler effect. Compared with OFDM, the orthogonal time-frequency-space modulation OTFS has better performance in a high-mobility wireless communication scene. The OTFS is a two-dimensional modulation scheme for performing modulation in the delay-doppler domain, and converts a double-dispersion channel into a channel that is approximately non-fading in the delay-doppler domain through a series of two-dimensional transformations. Two challenges facing OTFS systems are: how to accurately estimate the CSI, another is to obtain the CSI and then to perform received signal detection with a low complexity and high efficiency algorithm. The received signal detection is to detect a corresponding symbol corresponding to the transmitted symbol from the received signal, and if the complexity of the detection algorithm of the OTFS system is high, the detection of the received symbol by the whole system is slow and high time delay is caused, which is not favorable for the practicability of the actual system.

In recent years, with the development of big data, supercomputing, neuroscience and the like, artificial intelligence is rapidly developed, and some bottleneck problems in the development process of a neural network are solved, so that the realization of the neural network with high efficiency, high accuracy, high integration, low power consumption and low cost becomes possible, and the method plays an important role in many fields. Deep neural networks DNN have enjoyed tremendous success in the fields of image processing, natural language processing, and language recognition. DNN is a multilayer neural network structure formed by complex connection, and the multilayer structure has more expressive force and stronger fitting capability than a few neural network layer structures. DNN networks have found widespread use in various fields, and have also been successfully used in the communications field. Their use in the design of the physical layer of wireless communications is gaining increasing interest, and constellation design, transceiver design, coding design using automatic encoders, channel estimation, signal detection and demodulation are some of these areas. In particular, in the case of re-detection, the detection problem can also be regarded as a classification problem, and the DNN network has a strong classification capability, and therefore is very suitable for detecting the communication signal. In addition, with the continuous maturity of the neural network technology, conditions are created for the application of the neural network technology to high mobility devices such as high-speed rails and automobiles, and the problem of difficult detection of the OTFS system can be effectively solved.

Therefore, the detection of the OTFS signal is carried out by utilizing the deep neural network, the effective detection can be carried out by depending on the strong learning ability, the method not only can obtain lower complexity and higher performance than the traditional detection method, but also can not carry out channel estimation, thereby sending more data.

The existing OTFS detector includes traditional detection methods such as a minimum mean square error detector MMSE, a maximum likelihood detector ML, and various iterative detectors based on a message passing MP and its deformation, which are either high in complexity of a detection algorithm or poor in detection performance, and therefore, a low-complexity and high-efficiency detection algorithm is urgently needed. In addition, the conventional method needs to perform accurate channel estimation, which not only has a complex algorithm, but also causes additional overhead due to pilot frequency and guard interval.

In order to overcome the disadvantages of the conventional Detection methods, ashwitha Naikoti and A.Chockalingam in its published article "Low-complexity Delay-Doppler Symbol DNN for OTFS Signal Detection" (IEEE 93rd Vehicular Technology Conference 2021) mention one of the proposed Low-complexity OTFS Detection methods based on DNN, one is a FULL-DNN network, and the other is a Symbol DNN detector which performs Detection Symbol by Symbol. The symbol-based DNN detector adopts MN DNN networks to perform symbol-by-symbol detection, the number of output neurons of each neural network is determined by points mapped by a constellation diagram, so that the number of output neurons linearly increases along with the increase of MN, and the number of output neurons of the FULL-DNN network exponentially increases in the size of a transmission symbol vector, so that the complexity is obviously lower than that of the FULL-DNN, the performance of the detector is equivalent to that of the FULL-DNN, and the performance of the detector is approximate to that of maximum likelihood detection ML. In contrast, the detection performance of both networks is close, but the complexity of the symbol DNN is significantly lower. Although the second method reduces a large number of neurons, the method uses the data of the whole delay-doppler domain for detection, so that a large number of neurons are still required at the input end of the deep neural network, and a large number of neurons need to be increased in the hidden layer in proportion, so that more parameters need to be trained, and the detection speed of received symbols in the OTFS communication system is slow.

The patent document of the Beijing post and telecommunications university with the application number CN202010158335.1 proposes a signal detection method and a signal detection device of an OTFS system. The method comprises the steps of firstly establishing a corresponding factor graph, and establishing a neural network according to the factor graph, wherein the number of layers of a hidden layer of the neural network is the same as the iteration number of message transmission, and the hidden layer comprises a message computation neuron and a probability computation neuron. The message computation neurons correspond to nodes and/or edges of the factor graph; the probability calculation neuron is used for calculating data output by the neuron according to the signal detection performance parameters and the information, and calculating the probability of each modulation symbol in a transmission signal obtained after the transmission signal passes through a channel. The method utilizes the neural network for training to obtain optimized signal detection performance parameters, thereby improving the signal detection performance. But has the following disadvantages: the complexity of the detection method is related to the neural network and the iterative AMP algorithm used, when the number of symbols transmitted by each frame is large, the number of iterations is large, and the complexity of corresponding detection is also greatly increased, so that the method is not suitable for scenes with large total number of symbols of the transmitted frames, and the number of network layers is large, so that the whole neural network system is complex and a large number of parameters need to be trained.

Disclosure of Invention

The invention aims to provide a low-complexity OTFS system symbol detection method based on a deep neural network in order to improve the speed of detecting a received symbol by an OTFS communication system, simplify the detection complexity and improve the transmission efficiency.

To achieve the above object, the implementation scheme of the present invention comprises the following steps:

(1) Obtaining a training set;

1a) The OTFS system transmitting terminal sends the modulated signal, one-hot coding is carried out on the point on the constellation diagram of the transmitted signal, and the coded data is stored as a label;

1b) Receiving end of OTFS system receives data passing through channel, converting it to time delay-Doppler domain, and storing MN receiving symbols of the domain;

1c) Circulating the steps 1 a) and 1 b) for C times to obtain a training set, wherein C is determined according to the performance of the OTFS system;

(2) MN deep neural networks with completely identical structures are built, each network comprises an input layer, a hidden layer and an output layer, and all the layers are connected in a full-connection mode; marking MN networks to respectively correspond to MN receiving symbols in the OTFS system, wherein M and N respectively represent the total number of subcarriers and the total number of symbols of the OTFS system;

(3) Training each deep neural network using a training set:

3a) According to the corresponding relation between the time delay and the Doppler of the OTFS system, all data in the maximum time delay and the maximum Doppler range of the detection symbol corresponding to the network are used as input signals and input into an input layer of a neural network;

3b) Training the deep neural network by adopting a gradient descent algorithm to obtain data of an output layer;

3c) Adopting a minimum mean square error function as a loss function of the deep neural network, comparing the output layer data obtained in the step 3 b) with the label of the step 1 a), and judging whether the loss function is converged:

if yes, finishing training to obtain a trained deep neural network:

otherwise, returning to the step (3 b);

(4) A receiving end of the OTFS system receives a time domain signal sent by a transmitting end of the OTFS system, and carries out Wigner transformation on the time domain signal to obtain a signal Y [ n, m ] of a time-frequency domain;

(5) Carrying out Fourier transform (SFFT) on the signals in the time-frequency domain to obtain MN receiving symbols in the delay-Doppler domain;

(6) Adopting MN deep neural networks after training, carrying out symbol-by-symbol detection on received symbols:

6a) The MN symbols received in the step (5) correspond to the MN deep neural networks trained in the step (3), and the deep neural network corresponding to each received symbol is found;

6b) Taking all data in the maximum time delay and the maximum Doppler range of each received symbol as corresponding trained deep neural network input layer data to obtain output layer data of each neural network;

(6c) And obtaining an estimated value of the sending symbol according to the output layer data of each neural network, and completing the detection of the OTFS system symbol.

Compared with the prior art, the invention has the following advantages:

firstly, the invention utilizes the characteristic of the OTFS system in the delay-Doppler domain to select the data in the range of the maximum delay and the maximum Doppler domain of each symbol as the input end of the neural network for detection, rather than selecting all data for detection, so that the input data is reduced, the parameters of the neural network which need to be learned are reduced, and the detection complexity is reduced;

secondly, the invention avoids the difficulty of channel estimation and saves the expenses of pilot frequency and guard interval by utilizing the capability of extracting data characteristics of the deep neural network, thereby transmitting more effective data and improving the transmission efficiency.

Drawings

FIG. 1 is a flow chart of an implementation of the present invention;

FIG. 2 is a diagram of a deep neural network constructed in accordance with the present invention;

FIG. 3 is a data plot of maximum delay versus maximum Doppler domain range for the input layer of a neural network in accordance with the present invention;

fig. 4 is a simulation comparison diagram of the detection result of the OTFS system symbol according to the present invention and the existing detection method.

Detailed Description

The embodiments and effects of the present invention are further described in detail below with reference to the accompanying drawings:

referring to fig. 1, the implementation steps of this embodiment are as follows:

step 1, running an OTFS system by using MATLAB software to obtain a training set.

1.1 The OTFS system transmitting end transmits modulated signals, points on a constellation diagram are sequenced into 1,2, …, Q, … and Q according to the sequence from left to right and from top to bottom, for the Q-th point, the Q-th bit in the corresponding one-hot code is 1, and the rest are 0.

BPSK modulation is adopted in the embodiment, the number of points on a constellation diagram is 2, the position of-1 on the constellation diagram is encoded into binary data 10 by one-hot, the position of 1 is encoded into binary data 01 by one-hot, and the encoded data is stored as a tag;

1.2 OTFS system receiving end receives data passing through the channel, transforms it to the delay-doppler domain, and stores MN received symbols of the domain:

the BPSK signal is placed in an MXN time delay-Doppler domain, and after the BPSK signal is subjected to time domain signals of two-dimensional inverse octant Fourier transform ISFFT and Heisenberg transform, the signal-to-noise ratio is selected to be 10dB, the carrier center frequency is 4GHz, the subcarrier interval is 15KHz, and the BPSK signal is transmitted through a multipath channel;

after receiving the time domain signal, the receiving end obtains MN time delay-Doppler domain data through two-dimensional Fourier transform and Wigner transform, and MN receiving symbols are stored as training set data;

1.3 Step 1.1) and step 1.2) are circulated for C times, the number of training samples is selected through experiments, the principle of few to many training samples is adopted, a few examples are firstly started, then the use cases are gradually increased until a deep neural network tends to be stable, a training set is obtained, and C is determined according to the requirement of the error rate of the OTFS system;

in this example, M =16 and N =16 are respectively set to indicate the total number of subcarriers and the total number of carrier symbols of the OTFS system, the number of cycles C =30000, 8 transmission paths are used, and the positions of the paths are shown in table 1.

TABLE 1 location of the path at delay-Doppler

Route of travel	1	2	3	4	5	6	7	8
									Time delay (us)	0	4.16	8.32	12.48	16.64	20.8	24.96	29.12
Doppler (Hz)	0	0	938.5	938.5	938.5	1875	1875	1875
									Time delay offset	0	1	2	3	4	5	6	7
Doppler shift	0	0	1	1	1	2	2	2

And 2, constructing MN deep neural networks with completely identical structures by using a Tensorflow framework.

2.1 A three-layer structure including an input layer, a hidden layer and an output layer is provided, and all the layers are connected in a full-connection manner to build a deep neural network, as shown in fig. 2.

Determining the number of neurons of the input layer:

since the doppler is divided into positive doppler and negative doppler, (2 × k) is required _max + 1) neurons, and the input layer of the neural network is divided into (l) because the Doppler data is complex _max +1)×(2×k _max + 1) real and imaginary parts of data, thus the final determined number of input layer neurons: k =2 × (l) _max +1)×(2×k _max + 1), wherein l _max Is the index of the maximum delay of the OTFS system, k _max Is the index of the maximum doppler of the OTFS system;

the number of the neurons of the hidden layer is half of that of the input layer, namely the neurons of the hidden layer are (l) _max +1)×(2×k _max + 1) of;

the number of the neurons of the output layer is equal to the number Q =2 on the constellation diagram;

2.2 The OTFS is expressed in a matrix form, and the receiving and transmitting relations are as follows:

y＝Hx+n

wherein,

y denotes a received signal, H denotes a matrix, x denotes a transmitted signal, n denotes noise, and the (k + Nl) th element in the vector y satisfies the relationship y _k+Nl ＝y[k,l]In the same way, x and n also satisfy the relation, and elements in H are placed according to the corresponding relation of x and y.

2.3 Will be labeled 1,2, …, MN for MN deep neural networks with the same structure as 2.1) to correspond to the 1 st to MN elements of vector y in the OTFS system, respectively.

And 3, respectively training the MN deep neural networks by using the training sets.

Each network adopts the same training method, the example takes the p-th network as an example, and the training steps are as follows:

3.1 For OTFS transceiver system, let the delay tau on the ith path _i And Doppler upsilon _i Expressed as:

wherein l _i And k _i Indices representing delay taps and doppler taps, respectively, Δ f is the bandwidth of each subcarrier, T =1/Δ f;

3.2 Select the range of indices for maximum delay and maximum doppler:

in this example the maximum time delay tau _max =29.12 μ s, maximum doppler v _max =1.875KHz, with the indices of the corresponding maximum delay tap and maximum doppler tap being:

l _max ＝τ _max MΔf,k _max ＝υ _max NT

can find out _max ＝7，k _max =2, thenFor the p position y _p [k _p ,l _p ]，0≤k _p ≤N-1,0≤l _p ≤M-1,

For this position, the range of indices whose maximum delay is chosen is: l _p ≤L≤[l _p +l _max ] _M The index range of maximum Doppler is k _p -k _max ] _N ≤K≤[k _p +k _max ] _N In total of (l) _max +1)×(2×k _max + 1) data of which [.] _N Represents a cyclic shift of N.] _M Indicating that M is cyclically shifted, as indicated by the data within the dashed box in fig. 3, where the data within the solid box indicates the symbol for the p-th position;

3.3 For any training data in the training set, according to the corresponding relation of the OTFS system in the delay-Doppler domain, (l) in the index range described in step 3.2) _max +1)×(2×k _max + 1) separating the real part and the imaginary part of the data to form new training data;

3.4 Repeating the step 3.3) for C times to form a new C-dimensional training set;

3.5 Using a new training set, training the deep neural network by adopting a batch gradient descent method:

although the training method needs to consume a large amount of memory and has overlong training and learning time, the method can ensure that each updating is carried out towards the correct direction, and finally can ensure that the updating is converged at an extreme point, the performance of the method is superior to that of a random gradient descent algorithm and a small-batch gradient descent algorithm, and the method comprises the following specific implementation steps:

3.5.1 Adopting a random initialization method to carry out initial assignment on the training parameters omega and b;

3.5.2 C-dimensional training set data are sent to a network together, adam is adopted as an optimizer, and a minimum mean square error function is adopted as a loss function;

3.5.3 The deep neural network uses a Sigmoid function as an activation function of an output layer, and the remaining layers use Relu functions as activation functions to obtain data of the output layer of the network, wherein the activation function is used for introducing nonlinear factors, so that the learning capability of the network is stronger;

3.5.4 According to the training parameters omega and b at the current moment, calculating the gradient of the loss function at the current moment, and multiplying the learning rate by the gradient of the loss function to serve as the step length for updating the training parameters omega and b at the next time;

3.5.5 According to the output layer data and the label of the neural network, calculating the size of the loss function, and observing whether the loss function is converged:

if convergence, keeping the training parameters omega and b in the step 3.5.4) as a final training result;

if the loss function is not converged, the training parameters ω and b are updated again according to the step size in step 3.5.4) to advance in the direction of gradient descent, and then the step 3.5.2) is returned to perform retraining until the loss function is converged.

And 4, acquiring a signal of a time-frequency domain by a receiving end of the OTFS system.

4.1 At the transmitting end, the transmitted signal in the delay doppler domain is X [ k, l ], a time-frequency domain signal X [ n, m ] is obtained through two-dimensional symplectic inverse fourier transform (ISFFT), and the time-domain signal X (t) is obtained by performing Heisenberg transform on X [ n, m ], which are respectively expressed as follows:

in the formula g _tx (t) transmitting waveform pulses, the time domain signal being transmitted over a wireless channel;

4.2 Receiving end of OTFS system receives time domain signal y (t) sent by transmitting end:

h (tau, upsilon) represents channel information, tau represents a time delay index, and upsilon represents a Doppler index;

4.3 Carrying out Wigner transformation on the time domain signal Y (t) to obtain a signal Y [ n, m ] of a time-frequency domain;

Y[n,m]＝A _grx,y (t,f)| _{t＝nT,f＝mΔf}

wherein, A _grx,y (T, f) represents the cross-ambiguity function obtained by the matched filter at the receiving end, T represents the time index, f represents the frequency index, Δ f is the bandwidth of each subcarrier, T =1/Δ f represents the time interval, g _rx (t) is a received waveform pulse at the receiving end, g _r ^* _x (t) denotes a conjugate transpose, and j denotes an imaginary part.

And step 5, the receiving end obtains MN receiving symbols in the delay-Doppler domain.

Performing two-dimensional Fourier transform (SFFT) on the signals Y [ n, m ] in the time-frequency domain to obtain MN receiving symbols in the delay-Doppler domain:

wherein, y [ k, l]Denotes one received symbol with delay index k and doppler index l in the delay-doppler domain, k =0,1.. N-1, l =0,1.. M-1,P denotes the number of channel paths,

h _i representing the channel coefficient, v, on the ith path _i ，τ _i Respectively representing the delay and doppler, l, over the path _i And k _i Indexes, v [ k, l ], respectively representing delay taps and Doppler taps of the ith path]Representing the noise at the location of delay index k and doppler index l in the delay-doppler domain.

And 6, adopting the trained MN deep neural networks to detect the received symbols one by one.

6.1 MN receiving symbols in the delay-doppler domain are obtained to respectively correspond to the MN deep neural networks after training, and any one of the MN receiving symbols is detected by adopting the same method;

6.2 For the p-th received symbol y therein _p [k _p ,l _p ]Within the range of the dotted line in FIG. 3, (/) is selected _max +1)×(2×k _max + 1) position data, with a delay index of l _p ≤L≤[l _p +l _max ] _M Doppler index range of

[k _p -k _max ] _N ≤K≤[k _p +k _max ] _N ；

6.3 Separate the real part and imaginary part of the data in the range of the delay index and the range of the Doppler index to obtain

2×(l _max +1)×(2×k _max + 1) data, which is used as the input layer signal of the p-th deep neural network and is passed through the network _max +1)×(2×k _max + 1) obtaining the output of the neuron after hiding layer neurons and Q output layer neurons;

6.4 The Sigmoid function is adopted by the output layer as an activation function, the output range of each neuron is within the interval of 0,1, the data of all output neurons are traversed to obtain the maximum data,

6.5 The positions of the maximum data are marked as a, a =1,2, … … and Q, the points on the constellation diagram are sorted as 1,2, …, a, … and Q according to the sequence from left to right and from top to bottom, wherein the data corresponding to the point on the a-th constellation diagram is the estimated value of the transmitted symbol.

For the present example, Q =2 represents a point having a value of-1 on the constellation diagram when a =1, and represents a point having a value of 1 on the constellation diagram when a = 2.

The effect of the invention can be further illustrated by the following simulation experiment:

simulation conditions

The hardware platform of the simulation experiment is as follows: the processor is an Intel i3 8100CPU, the main frequency is 3.6GHz, and the memory is 8GB.

The software platform of the simulation experiment is as follows: windows 10 operating system, MATLAB R2020b and python3.6.

The channel type of the simulation experiment is a complex Gaussian channel, the channel path number is respectively selected to be 8, the subcarrier interval is 15KHz, the maximum Doppler is 1.875KHz, the maximum time delay is 29.12 mus, and the cycle number of the statistical bit error rate is 1000000.

The simulation and network parameters are shown in table 2.

TABLE 2 simulation and network parameters

Parameter(s)	Notation DNN	Low complexity DNN
			Channel path
	8	8
			Number of subcarriers (M)	16	16
Total number of symbols (N)	16	16
			Carrier center frequency	4GHz	4GHz
Modulation system	BPSK	BPSK
			Number of input neuron	512	80
Number of output neurons	2	2
			Output layer activation function	Softmax	Sigmoid
Remaining layer activation function	Relu	Relu
			Optimizer	Adam	Adam
Loss function	Binary-crossentropy	MMSE
			Training SNR	10dB	10dB
Training set size	30000	30000
			Number of training sessions	100	100

Secondly, simulation result analysis:

in the simulation experiment 1, under the above simulation conditions, 1000000 received symbols of the OTFS system are respectively detected by the present invention, the existing MMSE technique and symbol-DNN technique, so as to obtain the corresponding detection error rate, and the result is shown in fig. 4. Wherein the abscissa represents the signal-to-noise ratio of the transmitted symbol, and the unit is dB; the ordinate represents the error rate of symbol detection.

In the simulation experiment of the present invention, two prior arts refer to:

MMSE technique: ahmad Nimr et al, "Extended GFDM frame: OTFS and GFDM

Comparison，2018IEEE Global Communications Conference(GLOBECOM),2018,pp.1-6.”

The detection method of the MMSE detector in the OTFS system proposed in (1).

symbol-DNN technique: a method for detecting Symbol-DNN in OTFS system is proposed in the published paper "Low-complex Delay-Doppler Symbol DNN for OTFS Signal Detection" (IEEE 93rd Vehicular Technology Conference 2021) by Ashwitha Naikoti et al.

As can be seen from fig. 4, the performance of the present invention is significantly better than the conventional MMSE detection scheme and the complexity is lower than that of the MMSE detector, and compared with the existing symbol-DNN network, the performance is similar, but the complexity is reduced.

Simulation experiment 2, respectively calculating the number of parameters which need to be trained under the condition that MN =256 for the present invention and the existing symbol-DNN technology, and the result is shown in table 3:

table 3m =n =16 numbers of training parameters

As can be seen from table 3, the amount of training data of the present invention is reduced by 40 times compared to the prior art. The present invention utilizes the data of the maximum delay and the maximum doppler to detect, for each received symbol to be detected, the input end of the detector is all data in the range of the maximum delay index and the maximum doppler index of the symbol, and according to the characteristic that any symbol of the OTFS system does not exceed the maximum delay and the maximum doppler, the present invention utilizes the signal in the range to extract the features. Although the scheme contains a lot of useless information, the useless information can still be ignored by virtue of the strong learning capacity of the neural network, compared with the prior art of symbol-DNN, which uses all signals of the whole time delay-Doppler domain for detection, the useless information in the invention is far less than that of the technique of symbol-DNN, the parameters to be trained and the number of required neurons are also obviously less than that of the symbol-DNN network, and therefore, the complexity of the whole detection system is reduced.

Claims

1. The OTFS system symbol detection method based on the deep neural network is characterized by comprising the following steps

(1) Obtaining a training set;

(3) Training each deep neural network using a training set:

if so, finishing the training to obtain a trained deep neural network:

otherwise, returning to the step (3 b);

(5) Performing octave Fourier transform (SFFT) on the signals in the time-frequency domain to obtain MN receiving symbols in the delay-Doppler domain;

2. The method according to claim 1, wherein the one-hot encoding is performed on the points on the constellation diagram of the transmitted signal in step 1 a) as follows:

1a1) Determining the number Q of points on a constellation diagram, and taking the number Q as the bit number of one-hot coding;

1a2) The points on the constellation diagram are sorted into 1,2, …, Q, … and Q according to the sequence from left to right and from top to bottom, and for the Q-th point, the Q-th bit in the corresponding one-hot code is 1, and the rest are 0.

3. The method of claim 1, wherein the deep neural network is trained in step 3 b) using a gradient descent algorithm to obtain output layer data as follows:

3b1) Calculating the gradient of a loss function at the current moment according to the training parameters at the current moment;

3b2) Multiplying the step length by the gradient of the loss function to obtain the descending distance of the current position;

3b3) Setting a threshold value epsilon according to the bit error rate requirement of the OTFS system, and judging whether the gradient descending distance of all training parameters is smaller than epsilon:

if the training parameters are smaller than epsilon, all the current training parameters are the final output layer data;

otherwise, updating all training parameters and returning to the step (3 b 1).

4. The method according to claim 1, wherein the time-frequency domain signal Y [ n, m ] obtained in step (4) is represented as follows:

Y[n,m]＝A _grx,y (t,f)| _{t＝nT,f＝mΔf}

wherein A is _grx,y (T, f) represents the cross-ambiguity function obtained by the matched filter at the receiving end, T represents the time index, f represents the frequency index, Δ f is the bandwidth of each subcarrier, T =1/Δ f represents the time interval, g _rx (t) is a received waveform pulse at the receiving end,

denotes the conjugate transpose, j denotes the imaginary part, y (t) denotes the received time domain signal as:

y(t)＝∫ _υ ∫ _τ h(τ,υ)x(t-τ)e ^j2πυ(t-τ) dτdυ

h (tau, upsilon) represents channel information, tau represents a time delay index, upsilon represents a Doppler index, and x (t) represents a time domain transmitting signal of a transmitting terminal.

5. The method of claim 1, wherein the step (5) obtains MN received symbols in the delay-doppler domain as follows:

h _i representing the channel coefficients on the ith path, where v _i ，τ _i Respectively representing the delay and the Doppler on the path,/ _i And k _i Indices representing the delay tap and the doppler tap of the ith path, respectively, [.] _N Represents a cyclic shift of N.] _M Denotes cyclic shift of M, v [ k, l]Representing the noise at the location of delay index k and doppler index l in the delay-doppler domain.

6. The method of claim 1 wherein the step (6 c) of obtaining an estimate of the transmitted symbols from the output layer data of each neural network is

6c1) Traversing Q output layer data of the deep neural network to obtain the position of the maximum value, and marking as a;

6c2) The points on the constellation diagram are sorted into 1,2, … … and Q according to the sequence from left to right and from top to bottom, wherein the data corresponding to the point on the a-th constellation diagram is the estimated value of the transmitted symbol.