CN115514596B - OTFS communication receiver signal processing method and device based on convolutional neural network - Google Patents

Abstract

The present invention discloses a method and device for processing signals of an OTFS communication receiver based on a convolutional neural network. The method comprises: obtaining a training data set generated by an OTFS transmitter at the transmitting end of the OTFS communication system; training a deep convolutional neural network through the training data set to obtain a trained OTFS receiver signal processing model; using a signal received by the receiver as an input of the OTFS receiver signal processing model; and obtaining a processed received signal according to the output of the OTFS receiver signal processing model. The method can recover information with a lower bit error rate, recover the signal received by the receiver, and improve the reliability of wireless communication.

Description

OTFS communication receiver signal processing method and device based on convolutional neural network

Technical Field

The present invention relates to the field of wireless communication technology, and in particular to a convolutional neural network-based OTFS communication receiver signal processing method and device.

Background technique

With the large-scale construction and deployment of highways and high-speed railways, as well as the increasing popularity of autonomous driving technology. Internet of Vehicles (IoV), as one of the key application scenarios of the fifth generation (5G) mobile communication technology, has become an indispensable demand of users in the future. However, orthogonal frequency division multiplexing (OFDM), which is widely used in 5G mobile communication systems, is very sensitive to high Doppler shift effects, which makes OFDM's performance poor under fast time-varying channels and difficult to meet the growing needs of future IoV systems. In order to solve the low-latency and high-reliability communication problems of IoV systems in high-speed mobile scenarios. Orthogonal Time-Frequency-Space (OTFS) is a two-dimensional modulation technology suitable for dual-dispersion fading channels. At the same time, OTFS can be implemented on the basis of OFDM, that is, by adding additional pre-processing and post-processing modules to be compatible with the Long Term Evolution (LTE) architecture.

At present, edge computing based on deep learning plays an important role in resource scheduling and load balancing of the Internet of Vehicles. However, on the one hand, in edge computing based on the Internet of Vehicles, most of the research on deep learning remains at the network layer. On the other hand, for the research on physical layer communication of the Internet of Vehicles, deep learning is mostly used to optimize the performance of each communication module, but the local optimality of each module of the communication system is not the overall optimal performance of the receiver.

Therefore, people in this field are in urgent need of finding a new technical solution to solve the above problems.

Summary of the invention

In order to overcome the problems existing in the related art, the present invention discloses a method and device for processing signals of an OTFS communication receiver based on a convolutional neural network.

According to a first aspect of an embodiment disclosed in the present invention, there is provided an OTFS communication receiver signal processing method based on a convolutional neural network, the method comprising:

At the transmitting end of the OTFS communication system, a training data set generated by the OTFS transmitter is obtained;

The deep convolutional neural network is trained by using the training data set to obtain a trained OTFS receiver signal processing model;

The signal received by the receiver is used as the input of the OTFS receiver signal processing model;

According to the output of the OTFS receiver signal processing model, a processed received signal is obtained.

Optionally, the using the signal received by the receiver as an input of the OTFS receiver signal processing model includes:

The signal received by the receiver is mapped in a preset rectangular coordinate system to obtain the real part Re(r) and the imaginary part Im(r) of the IQ signal;

The real part Re(r) and the imaginary part Im(r) of the IQ signal are used as inputs of the OTFS receiver signal processing model.

Optionally, the method further includes:

The training data set is:

The loss function is determined as:

Among them, _NB is the number of samples contained in the mini-batch, _Tni is the true label of the i-th category of the n-th sample, and _Pni is the output probability of the i-th category of the n-th sample.

Optionally, after taking the signal received by the receiver as the input of the OTFS receiver signal processing model, the method includes:

Extracting shallow features in the input of the OTFS receiver signal processing model through a shallow feature extraction layer, wherein the shallow feature extraction layer includes three convolutional layers, a batch normalization layer and a ReLU activation layer;

The output of the shallow feature extraction layer is subjected to deep feature extraction through a backbone network, wherein the backbone network comprises a plurality of Bneck blocks.

Optionally, the convolutional neural network model is a MobileNetV2 lightweight one-dimensional convolutional neural network model;

The MobileNetV2 lightweight one-dimensional convolutional neural network model also includes: convolution layer, batch normalization, ReLu6 activation and global average pooling operations.

According to a second aspect of the disclosed embodiment of the present invention, there is provided an OTFS communication receiver signal processing device based on a convolutional neural network, the receiving card comprising: the device comprising:

The data set acquisition module, at the transmitting end of the OTFS communication system, acquires the training data set generated by the OTFS transmitter;

A model acquisition module, connected to the data set acquisition module, trains a deep convolutional neural network through the training data set to obtain a trained OTFS receiver signal processing model;

An input module, connected to the model acquisition module, takes the signal received by the receiver as the input of the OTFS receiver signal processing model;

The output module is connected to the input module and acquires the processed received signal according to the output of the OTFS receiver signal processing model.

Optionally, the input module includes:

A mapping unit maps the signal received by the receiver in a preset rectangular coordinate system to obtain the real part Re(r) and the imaginary part Im(r) of the IQ signal;

An input unit is connected to the mapping unit, and uses the real part Re(r) and the imaginary part Im(r) of the IQ signal as inputs of the OTFS receiver signal processing model.

Optionally, the training data set is:

The loss function is determined as:

Among them, _NB is the number of samples contained in the mini-batch, _Tni is the true label of the i-th category of the n-th sample, and _Pni is the output probability of the i-th category of the n-th sample.

Optionally, the device further comprises:

A shallow feature extraction module, which extracts shallow features in the input of the OTFS receiver signal processing model through a shallow feature extraction layer, wherein the shallow feature extraction layer includes three convolutional layers, a batch normalization layer and a ReLU activation layer;

The deep feature extraction module performs deep feature extraction on the output of the shallow feature extraction layer through a backbone network, wherein the backbone network includes a plurality of Bneck blocks.

Optionally, the convolutional neural network model is a MobileNetV2 lightweight one-dimensional convolutional neural network model;

The MobileNetV2 lightweight one-dimensional convolutional neural network model also includes: convolution layer, batch normalization, ReLu6 activation and global average pooling operations.

In summary, the present invention discloses a method and device for processing signals of an OTFS communication receiver based on a convolutional neural network, the method comprising: obtaining a training data set generated by an OTFS transmitter at the transmitting end of the OTFS communication system; training a deep convolutional neural network through the training data set to obtain a trained OTFS receiver signal processing model; using the signal received by the receiver as the input of the OTFS receiver signal processing model; and obtaining a processed received signal according to the output of the OTFS receiver signal processing model. The information can be recovered at a lower bit error rate, the signal received by the receiver can be recovered, and the reliability of wireless communication can be improved.

In addition, the novel signal processing method of the OTFS receiver in the disclosed embodiment of the present invention is different from using deep learning to optimize a certain information recovery module of the receiver, but optimizes each communication module of the receiver as a whole. A neural network is used to replace all modules of the receiving end (including carrier and symbol synchronization, channel estimation, equalization, demodulation, channel decoding, etc.) to complete the entire process of information recovery, avoiding the influence of imperfect channel state information (CSI) and cumulative errors caused by modular processing. In this way, the influence of factors such as inter-symbol interference (ISI) caused by multipath effects in wireless channels, inter-carrier interference (ICI) and Doppler interference (IDI) caused by Doppler frequency shift, and noise are overcome, ensuring low-latency and high-reliability wireless communication of the communication system in various complex scenarios.

Other features and advantages disclosed in the present invention will be described in detail in the following detailed description.

BRIEF DESCRIPTION OF THE DRAWINGS

The accompanying drawings are used to provide a further understanding of the present disclosure and constitute a part of the specification. Together with the following specific embodiments, they are used to explain the present disclosure but do not constitute a limitation of the present disclosure. In the accompanying drawings:

FIG1 is a flow chart of a signal processing method of an OTFS communication receiver based on a convolutional neural network according to an exemplary embodiment;

FIG2 is a schematic diagram of the mapping relationship between the delay-Doppler grid Γ and the time-frequency grid Λ;

FIG3 is a schematic diagram of the signal transformation relationship between the delay-Doppler domain and the time-frequency domain of the OTFS communication system;

FIG4 is a schematic diagram of a signal processing method model of an OTFS communication receiver;

Fig. 5 is a schematic diagram of a convolutional neural network;

FIG6 is a structural block diagram of an OTFS communication receiver signal processing device based on a convolutional neural network according to an exemplary embodiment;

FIG7 is a structural block diagram of an input module shown in FIG6 ;

FIG8 shows the bit error rate performance of the conventional OTFS communication receiver algorithm and the novel signal processing method;

Figure 9 shows the bit error rate performance of the conventional OTFS communication receiver in the EVA channel and the new signal processing method using the (7,4) Hamming code;

Figure 10 shows the bit error rate performance of the conventional OTFS communication receiver in the ETU channel and the new signal processing method using the (7,4) Hamming code;

FIG11 shows the bit error rate performance of the conventional OTFS communication receiver and the novel signal processing method using (7,4) Hamming code in the EVA channel;

FIG12 shows the performance of different implementation models of the novel signal processing method using (7,4) Hamming code in QPSK and 16QAM modulation modes.

Detailed ways

The specific embodiments disclosed in the present invention are described in detail below in conjunction with the accompanying drawings. It should be understood that the specific embodiments described herein are only used to illustrate and explain the present disclosure, and are not used to limit the present disclosure.

FIG. 1 is a flow chart of a signal processing method of an OTFS communication receiver based on a convolutional neural network according to an exemplary embodiment. As shown in FIG. 1 , the method includes:

Before introducing the OTFS communication receiver signal processing method based on a convolutional neural network in the disclosed embodiment of the present invention, the OTFS communication system and the deep convolutional neural network are first introduced.

The OTFS communication system includes a transmitter and a receiver. At the OTFS transmitter, the input signal x can obtain the matrix form of delay Doppler domain symbol X _DD through X _DD = vec ^-1 (x), and the time-frequency symbol X _TF = F _M X _DD F _N ^H is obtained by ISFFT transformation. The transmission symbol matrix S = G _tx F _M X _TF = G _tx X _DD F _N ^H is obtained by Heisenberg transformation. Finally, the MN×1 transmission signal vector is obtained by s = vec(S).

In the delay-Doppler channel, compared with the time-frequency domain channel response h(t,f), the delay-Doppler channel response h(τ,v) can better match the real physical channel through the delay-Doppler domain and intuitively display the different distances and relative speeds of multiple reflectors. As shown in Figure 4, compared with the time-frequency channel response extended to the entire time-frequency domain. Due to the limited number of reflectors superior to the channel, the channel response is more sparse in the delay-Doppler domain, and the channel response only exists in the grid range of (τ _max , ±v _max ), where τ _max is the maximum delay and v _max is the maximum Doppler frequency shift.

Due to the finite transmission paths and the associated delay and Doppler spread, the channel can be sparsely represented as:

Where P is the number of paths in the multipath channel, h _i , τ _i , and _vi are the channel gain, delay, and Doppler spread corresponding to the i-th path. The delay τ _i and Doppler spread value _vi can be converted from the indexes l _i , k _i in the delay-Doppler domain, where

The channel in matrix form can be expressed as:

H _eff is a MN×MN matrix, expressed as follows: where Π(Delay) is the permutation matrix (forward cyclic shift), (Doppler) is a diagonal matrix, where/>

The actual equivalent channel in matrix form for any pulse is

At the receiving end of the OTFS communication system, the received signal r(t) is obtained by superimposing the noise w(t) on the transmitted signal s(t) after passing through the Doppler channel h(τ,v)

r(t)＝∫∫h(τ,v)s(t-τ)e ^j2πv(t-τ) dτdv+w(t)

The received signal r = Hs + w is obtained through the channel. At the receiving end, the received signal is matrixed into R = vec ^-1 (r) to obtain the received signal matrix R. Then, the Wigner transform is used to obtain the time-frequency symbol Y _TF = F _M G _rx R. Then, the SFFT transform is used to obtain the delayed Doppler domain symbol Y _DD = F _M ^H Y _TF F _N = G _rx RF _N. Finally, Y _DD is vectorized to obtain our received signal

According to the equivalent simplified channel, the received signal can be expressed as

Among them, the mapping relationship between the delay-Doppler grid Γ and the time-frequency grid Λ is shown in Figure 2. On the delay-Doppler plane, N points are sampled along the Doppler axis at intervals of Doppler frequency shift resolution △v＝1/NT, and M points are sampled along the delay axis at intervals of delay spread Δτ＝1/MΔf. It is obtained that Γ＝{(k/NT,l/MΔf),k＝0,…,N-1,l＝0,…,M-1}. Through the two-dimensional symplectic finite Fourier inverse transform (ISFFT), the delay-Doppler plane is converted into the time-frequency domain grid Λ. In the time-frequency plane, N points are sampled along the time axis at intervals of T, and M points are sampled along the frequency axis at intervals of Δf＝1/T, and it is obtained that Λ＝{(nT,mΔf),n＝0,…,N-1,m＝0,…,M-1}. Assuming T = 1/Δf, in the time-frequency domain, the duration of the entire data packet is NT, and the occupied bandwidth is MΔf (where M is the number of subcarriers in the frequency domain; Δf is the subcarrier spacing; N is the number of time slots in the time domain, and T is the symbol duration).

CNN has the advantage of sharing convolution kernels, which allows the network to become deeper. By completing the gradual expression from shallow learning to deep learning of input data, more accurate features can be extracted to achieve better visual effects. However, some CNN models with excellent performance often have a deep network model structure and high model complexity, which is not conducive to mobile deployment. Among them, MobileNetV2 is a lightweight CNN model suitable for deployment on mobile or embedded devices. MobileNetV2 inherits the depthwise separable convolution of MobileNetV1. The reverse residual structure of MobileNetV2 also draws on the residual connection idea of ResNet. The difference is that MobileNetV2 first expands the dimension by mapping the input tensor from a low-dimensional space to a high-dimensional space through an expansion layer. Then, features are extracted using depthwise separable convolution. Finally, the output after the depthwise separable convolution is compressed through the projection layer, so that the data is mapped from a high-dimensional space to a low-dimensional space, and the network becomes smaller again. At the same time, since both the expansion layer and the projection layer contain learnable parameters, the entire network structure can learn how to better expand and compress data to achieve lightweight and better performance.

Although the Relu6 activation function model can make the model more robust under low-precision calculations, it will inevitably cause some feature loss. In order to reduce this information loss, only one batch normalization is added after the convolution of the projection layer instead of using Relu6.

In step 101, at the transmitting end of the OTFS communication system, a training data set generated by an OTFS transmitter is obtained.

For example, in the disclosed embodiment of the present invention, a lightweight CNN model is used to process the signal in the receiver to achieve reliable information recovery from the IQ signal waveform to the information bit stream. In the disclosed embodiment of the present invention, no module is optimized during the signal reception process, but a global optimization strategy is considered, and the intelligent processing of the OTFS communication reception signal is achieved through the 1D-Conv-MobileNetV2 structure.

It is understandable that before processing the signal received by the receiver, the convolutional neural network model needs to be trained to obtain the trained OTFS receiver signal processing model so as to process the signal received by the receiver according to the OTFS receiver signal processing model. When training the convolutional neural network model, it is necessary to first obtain a training data set.

Among them, the training data set is:

At the same time, the loss function needs to be determined as:

Among them, _NB is the number of samples contained in the mini-batch, _Tni is the true label of the i-th category of the n-th sample, and _Pni is the output probability of the i-th category of the n-th sample.

In step 102, a deep convolutional neural network is trained using the training data set to obtain a trained OTFS receiver signal processing model.

For example, the 1D-Conv-MobileNetV2 convolutional neural network model in the embodiment disclosed in the present invention is shown in FIG5 . It can be understood that the embodiment disclosed in the present invention has been architecturally designed and improved based on the original MobileNetV2.

In order to accurately extract the features of the OTFS communication system under the dual-dispersion channel, we first designed a shallow feature extraction layer, which consists of three convolutional layers, a batch normalization layer, and a ReLU activation layer. The feature extraction layer performs shallow feature extraction on the input. Then, using the dense connection idea in DenseNet, the output of the shallow feature extraction is connected in series with the input of the backbone network, and the deep features are further extracted through the backbone network. The linear activation is used to strengthen the feature transfer and improve the classification accuracy of the network. The backbone network consists of several Bneck blocks, where n is the number of Bneck repetitions. Then, the last part of 1D-Conv-MobileNetV2 consists of a convolutional layer, batch normalization, ReLu6 activation, and a global average pooling operation. Using average global pooling can reduce the number of network parameters while avoiding overfitting.

The goal of training is to optimize the network parameters based on the training data set generated by the transmitter so that the training model can achieve excellent performance and try to generalize to other data outside the training set. The training set of the new signal processing method is:

The loss function is the key to model training. The loss function used in this paper is cross entropy, which is defined as:

Where N _B is the number of samples in the mini-batch, T _ni is the true label of the ith category of the nth sample. P _ni is the output probability of the ith category of the nth sample. The optimization algorithm Adaptive Moment Estimation (Adam) used in this paper is an extension of the stochastic gradient descent (SGD) method. It combines the advantages of adaptive gradient (Adagrad) and root mean square back propagation (RMSProp) on the basis of SGD, and takes momentum into account. This means that Adam completes two optimizations: gradient sliding average and bias correction, solves sparse gradients, and maintains a relatively small amount of calculation, which can achieve good performance on non-stationary problems. Adam can not only further reduce parameter update jitter, but also balance the update speed of previous parameters, speed up convergence, and ensure final convergence.

The optimization process of the Adam algorithm is as follows: Calculate the exponentially weighted average of momentum

V _dθ = β ₁ V _dθ + (1 - β ₁ ) dθ

S _dθ = β ₂ S _dθ + (1 - β ₂ ) dθ ²

Update with RMSprop to get the first-order moment and second-order moment estimates respectively:

Last updated parameters

Among them, β ₁ ,β ₂ are the exponential decay rates of the instantaneous estimates, α is the learning rate, and ε is a constant to prevent division by zero. The weights W and bias b are based on The model parameters need to be updated in the network.

The training algorithm of the new signal processing method of the OTFS communication receiver is shown in Table 1:

Table 1 OTFS signal processing method training algorithm

In step 103, the signal received by the receiver is used as the input of the OTFS receiver signal processing model.

The difficulty of the OTFS receiver signal processing method lies in designing a data preprocessing method and a neural network structure suitable for complex OTFS communication signal processing. CNN usually uses two-dimensional black and white images or three-dimensional color images as input data. For communication problems, the form of input data is different from that of image data. The input data of this model is the real part Re(r) and imaginary part Im(r) of the received IQ signal. Specifically, the signal received by the receiver is mapped in a preset rectangular coordinate system to obtain the real part Re(r) and imaginary part Im(r) of the IQ signal; the real part Re(r) and imaginary part Im(r) of the IQ signal are used as the input of the OTFS receiver signal processing model.

Therefore, in the convolution layer of the network, this paper sets the channel of the convolution kernel to one dimension. The input of 1D-Conv-MobileNetV2 is the processed received signal, which can be expressed as:

In step 104, a processed received signal is obtained according to the output of the OTFS receiver signal processing model.

For example, the signal processing method of the OTFS communication receiver in the Internet of Vehicles is used to recover the signal information emitted by the traditional OTFS communication transmitter, in which the 1D-Conv-MobileNetV2 structure is used to replace the receiver of the traditional OTFS communication system to complete the entire process of receiving end information recovery. The purpose is to understand the complex relationship between the received signal and the transmitted information sequence, so as to restore the information under various non-ideal conditions as reliably as possible and improve the receiver's generalization ability to non-ideal conditions. The reliability of the wireless communication system is mainly reflected in the bit error rate. Therefore, the goal of the signal processing method design of the OTFS communication receiver is to minimize the bit error rate, which can be expressed as:

in is the information bit stream recovered by the signal processing method of the OTFS communication receiver, δ is the model parameter of the method, and F(·; δ) represents the function mapping from input to output of the method.

In addition, deploying deep learning algorithms to terminal devices requires consideration of memory and computing power requirements, so model complexity analysis is required. For new signal processing methods, the computational complexity of the general convolutional layer can be expressed as:

The computational complexity of the depth-separable convolution of MobileNetV2 can be expressed as:

In the formula, H _l , W _l represent the length and width of the feature map respectively, K _l represents the size and length of the convolution kernel, C _l-1 , C _l represent the number of input and output channels of the lth convolutional layer respectively.

The computational effort of the batch normalization layer and the ReLU layer is:

FIG6 is a structural block diagram of an OTFS communication receiver signal processing device based on a convolutional neural network according to an exemplary embodiment. As shown in FIG6 , the device 600 includes:

The data set acquisition module 610 acquires the training data set generated by the OTFS transmitter at the transmitting end of the OTFS communication system;

A model acquisition module 620, connected to the data set acquisition module 610, trains a deep convolutional neural network through the training data set to obtain a trained OTFS receiver signal processing model;

An input module 630, connected to the model acquisition module 620, uses the signal received by the receiver as an input of the OTFS receiver signal processing model;

The output module 640 is connected to the input module 630, and acquires the processed received signal according to the output of the OTFS receiver signal processing model.

FIG. 7 is a structural block diagram of an input module according to FIG. 6 . As shown in FIG. 7 , the input module 630 includes:

A mapping unit 631 maps the signal received by the receiver in a preset rectangular coordinate system to obtain the real part Re(r) and the imaginary part Im(r) of the IQ signal;

The input unit 632 is connected to the mapping unit 631, and uses the real part Re(r) and the imaginary part Im(r) of the IQ signal as inputs of the OTFS receiver signal processing model.

Optionally, the training dataset is:

The loss function is determined as:

Among them, _NB is the number of samples contained in the mini-batch, _Tni is the true label of the i-th category of the n-th sample, and _Pni is the output probability of the i-th category of the n-th sample.

Optionally, the device further comprises:

A shallow feature extraction module, which extracts shallow features in the input of the OTFS receiver signal processing model through a shallow feature extraction layer, wherein the shallow feature extraction layer includes three convolutional layers, a batch normalization layer and a ReLU activation layer;

The deep feature extraction module performs deep feature extraction on the output of the shallow feature extraction layer through a backbone network, and the backbone network includes several Bneck blocks.

Optionally, the convolutional neural network model is a MobileNetV2 lightweight one-dimensional convolutional neural network model;

The MobileNetV2 lightweight one-dimensional convolutional neural network model also includes: convolutional layer, batch normalization, ReLu6 activation and global average pooling operations.

The performance simulation of the OTFS communication receiver signal processing method based on convolutional neural network proposed in this invention is carried out:

(1) Simulation parameter settings

At the transmitter of the OTFS wireless communication system, the OTFS frame structure (M, N) is (4, 7), and the information bit stream is randomly generated. The carrier frequency f _c is 4 GHz, and the subcarrier spacing Δf is 15 KHz. Modulation uses binary phase shift keying (BPSK), quadrature phase shift keying (QPSK) and quadrature amplitude modulation (QAM). Channel coding uses (7, 4) Hamming coding, and the equalization algorithm uses the message passing algorithm (MP algorithm).

The training set, validation set and test set of the OTFS communication receiving signal processing method are generated by MATLAB simulation. During the model training process, the range of Eb/N0 is 0-8dB with an interval of 1dB. At each Eb/N0, the number of samples in the training set is 40,000 and the number of samples in the validation set is 20,000. In the test set, the sample capacity is 40,000. In order to verify the generalization ability of the model, for the Eb/N0 samples in the untrained test set, the value range of Eb/N0 during the test is 0-8dB with an interval of 0.5dB. The optimization algorithm uses Adam, the default learning rate α is 0.001, and the exponential decay rates β1 and β2 are 0.9 and 0.999 respectively. The initial learning rate is set to 0.001. During training, the mini-batch size is set to 256 and the maximum training period is 8.

1) Impact of additive white Gaussian noise channel:

Figure 8 shows the bit error rate performance of the traditional OTFS communication receiver algorithm and the new signal processing method. The receiver uses hard decision and maximum likelihood (ML) estimation for decoding, respectively. The traditional use of hard decision in the receiver refers to a decoding method that is not affected by any other factors except additive white Gaussian noise (AWGN). Since the equiprobability distribution of information bits simulates random generation, ML decision represents the optimal performance under ideal conditions. It can be seen that when Eb/N0 is 8dB, the bit error rate of the traditional OTFS communication receiver can reach ^10-5 in both BPSK and QPSK modulation modes. At the same time, the performance of the signal processing method is better than that of the traditional OTFS receiver. In the QPSK modulation mode, the proposed method can achieve a bit error rate close to ^10-6 when Eb/N0 is 7dB, and a bit error rate of ^10-6 when Eb/N0 is 8dB. When Eb/N0 is 7dB, the bit error rate of the proposed method in BPSK modulation mode can reach 10 ^-6 , and when Eb/N0 is 8dB, the bit error rate is 0.

The performance of this new signal processing method in terms of bit error rate is very close to the ideal ML decision, which is much better than the traditional hard decision method, and also shows that it has the potential to approach the performance limit. On the untrained Eb/N0, the new signal processing method also achieved performance close to that of ML decision, indicating that the method has good generalization ability for Eb/N0.

2) The impact of different vehicle channel conditions:

Figure 9 shows the bit error rate performance of the traditional OTFS communication receiver in the EVA channel and the new signal processing method using the (7,4) Hamming code. Under different terminal moving speeds, it can be seen that whether the traditional receiver adopts BPSK or QPSK modulation, when Eb/N0 is 8dB, the bit error rate value is between ^10-2 and ^10-3 . At a terminal moving speed of 350Kmph, the bit error rate performance of the traditional receiver is basically the same as that of 500Kmph. Under different terminal moving speeds, the performance of this algorithm is significantly better than that of the traditional algorithm. When Eb/N0 is 8dB, the bit error rates in the above cases are ^10-5 and ^10-6 . In the BPSK modulation mode, when Eb/N0 is 8dB, the bit error rate is close to ^10-6 when the terminal moving speed is 350Kmph.

Figure 10 shows the bit error rate performance of the traditional OTFS communication receiver in the ETU channel and the new signal processing method using the (7,4) Hamming code. As can be seen from Figure 10, whether the traditional receiver uses BPSK modulation or QPSK modulation, when Eb/N0 is 8dB, the new signal processing method can recover information with a lower bit error rate at different terminal moving speeds compared to the performance of the traditional algorithm. The performance of the method proposed in this paper is similar in the ETU channel and the EVA channel and verifies the stability and reliability of the method under different channel conditions.

3) Impact of different modulation modes:

The high-order modulation method used in actual communication systems can improve spectrum efficiency and anti-interference capabilities, but it has high requirements for signal detection. The modulations used are QPSK, 8QAM and 16QAM respectively. Figure 11 shows the bit error rate performance of the traditional OTFS communication receiver and the new signal processing method using (7,4) Hamming code under the EVA channel. The terminal moves at a speed of 500 kilometers per hour. When the moving speed is 500Kmph, when Eb/N0 is 8dB, the bit error rate performance of the traditional receiver is between ^10-1 and ^10-2 using different modulation methods. In the new signal processing method, when Eb/N0 is 8dB, the BER of 16QAM modulation reaches ^10-4 , the BER of 8QAM modulation is between ^10-4 and ^10-5 , and the BER of QPSK modulation is between ^10-5 and ^10-6 . The simulation results show that the new signal processing method is still better than the traditional receiver using high-order modulation under (7,4) Hamming code, which also proves that the method proposed in this paper still has high stability when using high-order modulation.

4) Effects of different signal processing methods on the model:

Figure 12 shows the performance of different implementation models of the new signal processing method using (7,4) in QPSK and 16QAM modulation modes. The vehicle network channel under Hamming code, the terminal moving speed is 500km/h. The improved network design is better than the original network model, which also reflects the potential of this new signal processing method in anti-interference. In QPSK modulation mode, when Eb/N0 is 8dB, the BER of the native ResNet model is close to ^10-4 , and the BER of the native DeneNet model, the native MobileNetV2 model and the proposed model are all between ^10-4 and ^10-5 . In 16QAM modulation mode, when Eb/N0 is 8dB, the BER of the native ResNet model reaches ^10-5 , and the BER of the native DeneNet model, the native MobileNetV2 model and the proposed model are all between ^10-5 and ^10-6 . By comparing the four different implementation methods, the method proposed in this paper obtains the best bit error rate performance.

We use different CNN models to implement intelligent signal processing methods and verify the impact of different CNN models on the reliability of the new signal processing method. From the above results, it can be clearly seen that 1D-Conv-MobileNetV2 not only reduces the amount of calculation in the information recovery process, but also achieves a lower bit error rate under different conditions, further improving the reliability of the receiver. The different implementations of this new signal processing method have a great impact on the reliability of the system, and the performance gap between different implementations is also obvious. Therefore, choosing a suitable CNN model and a reasonable optimization design have an important impact on the reliability of the new signal processing method.

In summary, the present invention discloses a method and device for processing signals of an OTFS communication receiver based on a convolutional neural network, the method comprising: obtaining a training data set generated by an OTFS transmitter at the transmitting end of the OTFS communication system; training a deep convolutional neural network through the training data set to obtain a trained OTFS receiver signal processing model; using the signal received by the receiver as the input of the OTFS receiver signal processing model; and obtaining a processed received signal according to the output of the OTFS receiver signal processing model. The information can be recovered at a lower bit error rate, the signal received by the receiver can be recovered, and the reliability of wireless communication can be improved.

In addition, the novel signal processing method of the OTFS receiver in the disclosed embodiment of the present invention is different from using deep learning to optimize a certain information recovery module of the receiver, but optimizes each communication module of the receiver as a whole. A neural network is used to replace all modules of the receiving end (including carrier and symbol synchronization, channel estimation, equalization, demodulation, channel decoding, etc.) to complete the entire process of information recovery, avoiding the influence of imperfect channel state information (CSI) and cumulative errors caused by modular processing. In this way, the influence of factors such as inter-symbol interference (ISI) caused by multipath effects in wireless channels, inter-carrier interference (ICI) and Doppler interference (IDI) caused by Doppler frequency shift, and noise are overcome, ensuring low-latency and high-reliability wireless communication of the communication system in various complex scenarios.

The preferred embodiments of the present disclosure are described in detail above in conjunction with the accompanying drawings; however, the present disclosure is not limited to the specific details in the above embodiments. Within the technical concept of the present disclosure, a variety of simple modifications can be made to the technical solution of the present disclosure, and these simple modifications all fall within the protection scope of the present disclosure.

It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. In order to avoid unnecessary repetition, the present disclosure will not further describe various possible combinations.

In addition, various embodiments of the present disclosure may be arbitrarily combined, and as long as they do not violate the concept of the present disclosure, they should also be regarded as the contents disclosed by the present disclosure.

Claims (4)

1. A convolutional neural network-based OTFS communication receiver signal processing method, characterized in that the method comprises: At the transmitting end of the OTFS communication system, a training data set generated by the OTFS transmitter is obtained; The deep convolutional neural network is trained by using the training data set to obtain a trained OTFS receiver signal processing model; The signal received by the receiver is used as the input of the OTFS receiver signal processing model; Acquire a processed received signal according to an output of the OTFS receiver signal processing model; The method further includes: the training data set is: , The loss function is determined as: , where /> is the number of samples contained in the small batch, /> For the first/> The first /> of the samples The true labels on the categories, /> It is the first/> The first /> of the samples Output probability of each category; The training method of the OTFS receiver signal processing model includes: inputting a training data set Maximum number of iterations/> , instantaneous estimate/> and learning rate/> ; Randomly initialize network parameters; Loop through the following steps: /> ; From the training set/> Randomly select from /> samples; calculate the loss function according to the formula; update the network parameters according to the Adam optimizer; repeat the above steps until the function mapping from input to output is obtained according to the minimization error training/> ; After the signal received by the receiver is used as the input of the OTFS receiver signal processing model, the method includes: extracting shallow features in the input of the OTFS receiver signal processing model through a shallow feature extraction layer, wherein the shallow feature extraction layer includes three convolutional layers, a batch normalization layer and a ReLU activation layer; performing deep feature extraction on the output of the shallow feature extraction layer through a backbone network, and the backbone network includes several Bneck blocks; the convolutional neural network model is a MobileNetV2 lightweight one-dimensional convolutional neural network model; the MobileNetV2 lightweight one-dimensional convolutional neural network model also includes: convolutional layer, batch normalization, ReLu6 activation and global average pooling operations. 2. The OTFS communication receiver signal processing method based on convolutional neural network according to claim 1, characterized in that the signal received by the receiver is used as the input of the OTFS receiver signal processing model, comprising: The signal received by the receiver is mapped in a preset rectangular coordinate system to obtain the real part Re(r) and the imaginary part Im(r) of the IQ signal; The real part Re(r) and the imaginary part Im(r) of the IQ signal are used as inputs of the OTFS receiver signal processing model. 3. A OTFS communication receiver signal processing device based on convolutional neural network, characterized in that the device comprises: The data set acquisition module, at the transmitting end of the OTFS communication system, acquires the training data set generated by the OTFS transmitter; A model acquisition module, connected to the data set acquisition module, trains a deep convolutional neural network through the training data set to obtain a trained OTFS receiver signal processing model; An input module, connected to the model acquisition module, takes the signal received by the receiver as the input of the OTFS receiver signal processing model; An output module, connected to the input module, for acquiring a processed received signal according to an output of the OTFS receiver signal processing model; The training data set is: The loss function is determined as: , where /> is the number of samples contained in the small batch, /> For the first/> The first /> of the samples The true labels on the categories, /> It is the first/> The first /> of the samples Output probability of each category; The training process of the OTFS receiver signal processing model includes: inputting a training data set Maximum number of iterations/> , instantaneous estimate/> and learning rate/> ; Randomly initialize network parameters; Loop through the following steps: /> ; From the training set/> Randomly select from /> samples; calculate the loss function according to the formula; update the network parameters according to the Adam optimizer; repeat the above steps until the function mapping from input to output is obtained according to the minimization error training/> ; The device also includes: a shallow feature extraction module, which extracts shallow features in the input of the OTFS receiver signal processing model through a shallow feature extraction layer, wherein the shallow feature extraction layer includes three convolutional layers, a batch normalization layer and a ReLU activation layer; a deep feature extraction module, which performs deep feature extraction on the output of the shallow feature extraction layer through a backbone network, and the backbone network includes several Bneck blocks; the convolutional neural network model is a MobileNetV2 lightweight one-dimensional convolutional neural network model; the MobileNetV2 lightweight one-dimensional convolutional neural network model also includes: convolutional layer, batch normalization, ReLu6 activation and global average pooling operations. 4. The OTFS communication receiver signal processing device based on convolutional neural network according to claim 3, characterized in that the input module comprises: A mapping unit maps the signal received by the receiver in a preset rectangular coordinate system to obtain the real part Re(r) and the imaginary part Im(r) of the IQ signal; An input unit is connected to the mapping unit, and uses the real part Re(r) and the imaginary part Im(r) of the IQ signal as inputs of the OTFS receiver signal processing model.