Channel estimation method of OTFS system in Internet of Vehicles based on improved convolutional neural network

Abstract

The channel estimation method of the OTFS system in the Internet of Vehicles based on the improved convolutional neural network belongs to the field of wireless communication technology. The method includes three steps: first, the OTFS wireless communication system in the Internet of Vehicles generates a delay Doppler (DD) domain data set, and uses the orthogonal matching pursuit algorithm to perform channel pre-estimation to obtain a noisy pre-estimated channel matrix; secondly, on the basis of the convolutional neural network, a mixed hole convolution and residual path are introduced, and the pre-estimated channel matrix and the actual channel matrix are input into the network for offline training; finally, under different signal-to-noise ratios, the pre-estimated channel matrix is ​​input into the trained network, and the final channel estimation matrix is ​​output. The present invention combines the OMP algorithm with the improved convolutional neural network, models the channel estimation problem as a denoising problem, and performs fast DD domain channel estimation on the OTFS wireless communication system in the Internet of Vehicles, which can significantly improve the accuracy of channel estimation and has good robustness to different moving speeds.

Description

Technical Field

[0001] The invention relates to an OTFS system channel estimation method in a vehicle networking system based on an improved convolutional neural network, and belongs to the technical field of wireless channel estimation. Background Art

[0002] Orthogonal Time-Frequency Space (OTFS) uses a new type of carrier in the delay-Doppler domain to achieve multiple transmission of QAM symbols. OTFS can make the transmission rate close to the channel capacity while achieving the optimal trade-off between performance and complexity, and effectively resist the negative effects of Doppler effect and multipath effect. Considering the importance of accurate channel estimation in the high-speed mobile environment of the Internet of Vehicles, the OTFS system can effectively balance the Doppler frequency shift by converting the channel with drastic changes in the time-frequency domain into a stable channel in the delay-Doppler domain. For the channel estimation problem of this system, traditional pilot-based channel estimation algorithms are generally used, such as least squares (LS) and orthogonal matching pursuit (OMP) algorithms. However, traditional channel estimation is greatly affected by noise signals, and the complexity increases sharply with the increase of the number of carriers. Deep learning, as an emerging technology in the field of artificial intelligence, especially convolutional neural networks, has been widely used in various aspects of wireless communication in recent years, such as channel estimation and signal detection. The improved convolutional neural network introduces the residual path on the basis of the convolutional neural network, and uses the normalized mean square error (NMSE) as the loss function, and models the channel estimation problem as a noise elimination problem, which can significantly reduce the impact of noise and has good robustness against the Doppler frequency shift caused by different speeds in the Internet of Vehicles. Compared with the traditional neural network method, it solves the problem of gradient disappearance while maintaining a low model complexity, and shows good application prospects in the field of wireless communications. The hybrid void convolution can obtain a larger receptive field, so that each convolution output contains a larger range of information, and at the same time, it can obtain more dense data features while reducing the amount of calculation, and thus more accurately extract noise features.

[0003] Qingyu Li et al. (see Q.Li, Y.Gong, F.Meng, Z.Li, L.Miao and Z.Xu, "ResidualLearning based Channel Estimation for OTFS system," 2022IEEE / CIC InternationalConference on Communications in China.) conducted channel estimation for the OTFS communication system based on deep learning, and proposed a delay-Doppler domain OTFS channel estimation technology based on model-driven deep learning (DL). The channel estimation results were further processed using a deep residual learning network (ResNet). The simulation results show that this method can achieve higher accuracy than traditional channel estimation methods. However, the improvement of channel estimation accuracy under low signal-to-noise ratio conditions is still lacking, and the model performance under large noise interference is not explored, and the model converges slowly. Summary of the invention

[0004] The traditional channel estimation algorithm based on pilot and OMP has low channel estimation accuracy when the noise interference is large, and the complexity increases sharply with the increase of carriers. In order to overcome this problem, the present invention proposes a channel estimation method for OTFS communication system based on improved convolutional neural network, which can eliminate noise under low signal-to-noise ratio conditions, thereby significantly improving the accuracy of channel estimation.

[0005] The technical solution of the present invention is as follows:

[0006] A channel estimation method for an OTFS system in an Internet of Vehicles based on an improved convolutional neural network is implemented by an OTFS wireless communication system, which includes a transmitter, a time domain channel and a receiver. The transmitter is a single-antenna user with an OTFS modulation module, and the OTFS modulation process includes an inverse symplectic Fourier transform and a Heisenberg transform; the number of sparse paths in the time domain channel is P; the receiver is a four-antenna receiver, which includes an OTFS demodulation module and a channel estimation module, and the OTFS demodulation process includes a symplectic Fourier transform and a Wigner transform; the OTFS wireless communication system can convert a channel with drastic changes in the time and frequency domain into a delay Doppler (DD) The stable channel in the domain can effectively suppress the Doppler frequency shift. The implementation process of the channel estimation module is as follows: first, the OTFS wireless communication system generates a data set, and uses the OMP algorithm to pre-estimate the channel to obtain a noisy pre-estimation matrix; secondly, the hybrid void convolution and residual path are introduced to build an improved convolutional neural network, and the noisy channel estimation matrix and the actual channel matrix are input into the network, and the data-driven method is used for offline training; finally, under different signal-to-noise ratios, the test data is input into the trained improved convolutional neural network, the final channel estimation matrix is ​​output, and the channel estimation performance is evaluated. The specific steps are as follows:

[0007] 1) OTFS wireless communication system simulation generates data sets and preprocesses the data:

[0008] First, a quadrature amplitude modulation (QAM) symbol frame of length M×N is arranged into a two-dimensional data matrix in represents an M×N-dimensional complex number set, where M and N are the number of units in the delay dimension and Doppler dimension respectively. An OTFS symbol frame is mapped to the delay-Doppler domain grid to obtain the signal x(k,l) in the delay-Doppler domain, where k∈{0,1,...,N-1}, l∈{0,1,...,M-1}; then, the time-frequency domain signal is obtained by the inverse symplectic Fourier transform (ISFFT) Where n and m represent the time domain index position and frequency domain index position of the time-frequency domain signal respectively; X[n,m] is then transformed into the time domain signal s(t) through Heisenberg transformation, which is expressed as Where t represents the sampling time of the time domain signal, g tx (t) is the time domain transmission pulse, T and Δf are the minimum sampling interval and minimum sampling frequency in the time-frequency domain respectively; then, s(t) is transmitted through the time domain channel, and the time domain signal r(t) is obtained at the receiving end, and r(t) = ∫∫h(τ,ν)g tx (t-τ)e j2πνtdτdν+v(t), where v(t) represents additive white Gaussian noise in the time domain, h(τ,ν) represents the channel impulse response with P sparse paths, τ and v represent the path delay and Doppler frequency shift in the delayed Doppler domain, respectively. r(t) is demodulated by Wigner transform at the receiving end to obtain the time-frequency domain output signal Y[n,m]=∫g rx (t-τ)r(t)e -j2π(t-τ) dt, where g rx (t) represents the received pulse in the time domain; finally, the delayed Doppler domain output signal is obtained by sigmoid Fourier transform (SFFT) The delay-Doppler domain channel estimation of the OTFS communication system requires inserting a pilot symbol x into the transmitted signal. p To pre-estimate the delay Doppler domain channel, the delay Doppler domain signal after inserting the pilot can be expressed as where x d Represents data symbol, (k p ,l p ) represents the position of the pilot symbol in the delay domain and Doppler domain, k max and l max denote the maximum delay and maximum Doppler spread in the delay-Doppler domain, respectively. p -2k max ,k p +2k max ] and l∈[l p -l max ,l p +l max ] is set to zero as the protection interval of the pilot symbol; the OMP algorithm is used for channel pre-estimation, the input pilot matrix is ​​represented as S, the delayed Doppler domain received signal is represented as Y, the starting residual r of the OMP algorithm is set to 1, the number of iterations is set to i, and the input pilot matrix S of the current iteration number i-1 is calculated i-1 All column vectors of the current residual r i-1 The product of , select the column vector s with the largest absolute value of the product i , add it to the pilot matrix of the next iteration i, denoted as S i =[S i-1 ,s i ], and the least squares method is used to obtain the channel estimation matrix of the i-th iteration The symbol · represents the norm matrix, that is, Smallest As And update the residual Update the current number of iterations i = i + 1. If the relationship i>P is satisfied, stop the iteration and update the last obtained As the final estimated result Otherwise, continue to execute the above OMP algorithm operation;

[0009] The dataset is generated by the OTFS wireless communication system and its size is 4×10 5 , 75% of the data is used for network training and 25% of the data is used for testing;

[0010] 2) Based on the convolutional neural network, an improved convolutional neural network model is built to estimate the channel matrix The actual channel matrix H is input into the network, the normalized mean square error (NMSE) is used as the loss function, and the data-driven method is used for offline training;

[0011] The mixed hole convolution and residual path are introduced into the convolutional neural network structure, and an improved convolutional neural network is built for noise elimination. The network includes two noise elimination modules, each of which contains three convolutional layers and a residual path. In view of the problems of resolution reduction and local information loss in convolutional neural networks, mixed hole convolution can be introduced in the convolutional layer to solve the above problems by reducing the image size and expanding the receptive field. The key to mixed hole convolution lies in the setting of the expansion coefficient. For the 3×3 convolution kernel of the network, when the expansion coefficient is 1, the convolution kernel size remains unchanged. When the expansion coefficient is 2, the convolution kernel hole increases by 1, and its size becomes 5×5. Therefore, when the expansion coefficients of the two superimposed convolutional layers are set to 1 and 2 respectively, the receptive field can be increased. However, if the expansion coefficient is increased multiple times, the kernel function will be discontinuous, resulting in loss of information continuity. In view of this situation, the expansion coefficient set when multiple convolutional layers are superimposed in the neural network should not be a common divisor greater than 1, thereby forming a sawtooth cyclic structure. The specific training steps of the improved convolutional neural network are as follows:

[0012] ① The channel pre-estimation matrix obtained by the OMP algorithm As input, the actual channel matrix H is used as the label;

[0013] ② Then, three convolution layers with a convolution kernel size of 3×3 are passed through, and the expansion coefficients of these three convolution layers are set to 1, 2, and 5 respectively, forming a sawtooth cyclic structure. The improved convolutional neural network distinguishes The difference between H and realizes the noise characteristic Extraction of

[0014] ③ Through the residual path and Subtract, the channel estimation result output by the output layer is The estimation result eliminates noise interference and is closer to the actual channel matrix H;

[0015] ④ The above steps ②③ can be regarded as the operations contained in the first noise elimination module. As the input of the second noise removal module, the actual channel matrix H is used as the label, and steps ②③ are repeated to obtain the channel estimation matrix of the improved convolutional neural network.

[0016] 3) Generate test data at different signal-to-noise ratios, input the trained improved convolutional neural network, output the estimated channel matrix, and further compare it with the actual channel matrix to evaluate the channel estimation performance:

[0017] After the channel estimation network training is completed, online deployment is realized. First, the OTFS communication system generates 4×10 4 The simulation data is estimated by OMP channel estimation algorithm. Then, the noise is eliminated through the channel estimation network based on the improved convolutional neural network, and finally a high-accuracy channel estimation matrix is ​​output. And compare it with the actual channel matrix H to evaluate the channel estimation performance.

[0018] The OTFS is the abbreviation of Orthogonal time frequency space, which means orthogonal time frequency space.

[0019] The DD is the abbreviation of Delay-Doppler, which means delayed Doppler.

[0020] The CNN is the abbreviation of Convolutional Neural Network, which means convolutional neural network.

[0021] The OMP is the abbreviation of Orthogonal Matching Pursuit, which means orthogonal matching pursuit.

[0022] The present invention proposes an uplink channel estimation method for an OTFS communication system based on an improved convolutional neural network. When the channel state is sparse, the OMP estimation algorithm is combined with the improved convolutional neural network to estimate the channel of the OTFS communication system, which can significantly improve the accuracy of channel estimation. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 It is a schematic diagram of the system framework of the method of the present invention.

[0024] Figure 2 It is a schematic diagram of the channel estimation network structure of the method of the present invention.

[0025] Figure 3The figure is a comparison of the normalized mean square error performance of the method of the present invention and traditional channel estimation methods such as pilot and orthogonal matching pursuit (OMP) under the configuration of carrier number M=128, subcarrier number N=8, and Fourier transform number FFT=512. Figure 3 It can be seen that the normalized mean square error of this method is much smaller than that of the pilot and OMP algorithms under low signal-to-noise ratio conditions. DETAILED DESCRIPTION

[0026] The present invention will be further described below with reference to the accompanying drawings and embodiments, but is not limited thereto.

[0027] Example:

[0028] A channel estimation method for an OTFS system in an Internet of Vehicles based on an improved convolutional neural network is implemented by an OTFS wireless communication system, which includes a transmitter, a time domain channel and a receiver. The transmitter is a single-antenna user with an OTFS modulation module, and the OTFS modulation process includes an inverse symplectic Fourier transform and a Heisenberg transform; the number of sparse paths in the time domain channel is P; the receiver is a four-antenna receiver, which includes an OTFS demodulation module and a channel estimation module, and the OTFS demodulation process includes a symplectic Fourier transform and a Wigner transform; the OTFS wireless communication system can convert a channel with drastic changes in the time and frequency domain into a delay Doppler (DD) The stable channel in the domain can effectively suppress the Doppler frequency shift. The implementation process of the channel estimation module is as follows: first, the OTFS wireless communication system generates a data set, and uses the OMP algorithm to pre-estimate the channel to obtain a noisy pre-estimation matrix; secondly, the hybrid void convolution and residual path are introduced to build an improved convolutional neural network, and the noisy channel estimation matrix and the actual channel matrix are input into the network, and the data-driven method is used for offline training; finally, under different signal-to-noise ratios, the test data is input into the trained improved convolutional neural network, the final channel estimation matrix is ​​output, and the channel estimation performance is evaluated. The specific steps are as follows:

[0029] 1) OTFS wireless communication system simulation generates data sets and preprocesses the data:

[0030] First, a quadrature amplitude modulation (QAM) symbol frame of length M×N is arranged into a two-dimensional data matrix in represents an M×N-dimensional complex number set, where M and N are the number of units in the delay dimension and Doppler dimension respectively. An OTFS symbol frame is mapped to the delay-Doppler domain grid to obtain the signal x(k,l) in the delay-Doppler domain, where k∈{0,1,...,N-1}, l∈{0,1,...,M-1}; then, the time-frequency domain signal is obtained by the inverse symplectic Fourier transform (ISFFT) Where n and m represent the time domain index position and frequency domain index position of the time-frequency domain signal respectively; X[n,m] is then transformed into the time domain signal s(t) through Heisenberg transformation, which is expressed as Where t represents the sampling time of the time domain signal, g tx (t) is the time domain transmission pulse, T and Δf are the minimum sampling interval and minimum sampling frequency in the time-frequency domain respectively; then, s(t) is transmitted through the time domain channel, and the time domain signal r(t) is obtained at the receiving end, and r(t) = ∫∫h(τ,ν)g tx (t-τ)e j2πνt dτdν+v(t), where v(t) represents additive white Gaussian noise in the time domain, h(τ,ν) represents the channel impulse response with P sparse paths, τ and v represent the path delay and Doppler frequency shift in the delayed Doppler domain, respectively. r(t) is demodulated by Wigner transform at the receiving end to obtain the time-frequency domain output signal Y[n,m]=∫g rx (t-τ)r(t)e -j2π(t-τ) dt, where g rx (t) represents the received pulse in the time domain; finally, the delayed Doppler domain output signal is obtained by sigmoid Fourier transform (SFFT) The delay-Doppler domain channel estimation of the OTFS communication system requires inserting a pilot symbol x into the transmitted signal. p To pre-estimate the delay Doppler domain channel, the delay Doppler domain signal after inserting the pilot can be expressed as where x d Represents data symbol, (k p ,l p ) represents the position of the pilot symbol in the delay domain and Doppler domain, k max and l max denote the maximum delay and maximum Doppler spread in the delay-Doppler domain, respectively. p -2k max ,k p +2k max ] and l∈[l p -l max ,l p +l max ] is set to zero as the protection interval of the pilot symbol; the OMP algorithm is used for channel pre-estimation, the input pilot matrix is ​​represented as S, the delayed Doppler domain received signal is represented as Y, the starting residual r of the OMP algorithm is set to 1, the number of iterations is set to i, and the input pilot matrix S of the current iteration number i-1 is calculated i-1 All column vectors of the current residual r i-1 The product of , select the column vector s with the largest absolute value of the product i , add it to the pilot matrix of the next iteration i, denoted as S i =[Si-1 ,s i ], and the least squares method is used to obtain the channel estimation matrix of the i-th iteration The symbol ||·|| indicates taking the norm matrix, that is, Smallest As And update the residual Update the current number of iterations i = i + 1. If the relationship i>P is satisfied, stop the iteration and update the last obtained As the final estimated result Otherwise, continue to execute the above OMP algorithm operation;

[0031] The dataset is generated by the OTFS wireless communication system and its size is 4×10 5 , 75% of the data is used for network training and 25% of the data is used for testing;

[0032] 2) Based on the convolutional neural network, an improved convolutional neural network model is built to estimate the channel matrix The actual channel matrix H is input into the network, the normalized mean square error (NMSE) is used as the loss function, and the data-driven method is used for offline training;

[0033] The mixed hole convolution and residual path are introduced into the convolutional neural network structure, and an improved convolutional neural network is built for noise elimination. The network includes two noise elimination modules, each of which contains three convolutional layers and a residual path. In view of the problems of resolution reduction and local information loss in convolutional neural networks, mixed hole convolution can be introduced in the convolutional layer to solve the above problems by reducing the image size and expanding the receptive field. The key to mixed hole convolution lies in the setting of the expansion coefficient. For the 3×3 convolution kernel of the network, when the expansion coefficient is 1, the convolution kernel size remains unchanged. When the expansion coefficient is 2, the convolution kernel hole increases by 1, and its size becomes 5×5. Therefore, when the expansion coefficients of the two superimposed convolutional layers are set to 1 and 2 respectively, the receptive field can be increased. However, if the expansion coefficient is increased multiple times, the kernel function will be discontinuous, resulting in loss of information continuity. In view of this situation, the expansion coefficient set when multiple convolutional layers are superimposed in the neural network should not be a common divisor greater than 1, thereby forming a sawtooth cyclic structure. The specific training steps of the improved convolutional neural network are as follows:

[0034] ① The channel estimation matrix obtained by the OMP algorithm As input, the actual channel matrix H is used as the label;

[0035] ② Then, three convolution layers with a convolution kernel size of 3×3 are passed through, and the expansion coefficients of these three convolution layers are set to 1, 2, and 5 respectively, forming a sawtooth cyclic structure. The improved convolutional neural network distinguishes The difference between H and realizes the noise characteristic Extraction of

[0036] ③ Through the residual path and Subtract, the channel estimation result output by the output layer is The estimation result eliminates noise interference and is closer to the actual channel matrix H;

[0037] ④ The above steps ②③ can be regarded as the operations contained in the first noise elimination module. As the input of the second noise removal module, the actual channel matrix H is used as the label, and steps ②③ are repeated to obtain the channel estimation matrix of the improved convolutional neural network.

[0038] 3) Generate test data at different signal-to-noise ratios, input the trained improved convolutional neural network, output the estimated channel matrix, and further compare it with the actual channel matrix to evaluate the channel estimation performance:

[0039] After the channel estimation network training is completed, online deployment is realized. First, the OTFS communication system generates 4×10 4 The simulation data is estimated by OMP channel estimation algorithm. Then, the noise is eliminated through the channel estimation network based on the improved convolutional neural network, and finally a high-accuracy channel estimation matrix is ​​output. And compare it with the actual channel matrix H to evaluate the channel estimation performance.

Claims

1. A channel estimation method for an OTFS system in an Internet of Vehicles based on an improved convolutional neural network is implemented by an OTFS wireless communication system, which includes a transmitter, a time domain channel and a receiver. The transmitter is a single-antenna user with an OTFS modulation module, and the OTFS modulation process includes an inverse sigmoid Fourier transform and a Heisenberg transform; the number of sparse paths in the time domain channel is P; the receiver is a four-antenna receiver, which includes an OTFS demodulation module and a channel estimation module, and the OTFS demodulation process includes a forward sigmoid Fourier transform and a Wigner transform; the OTFS wireless communication system can convert a channel with drastic changes in the time and frequency domain into a stable channel in the delayed Doppler domain, thereby effectively suppressing the Doppler frequency shift. The implementation process of the channel estimation module is as follows: first, a data set is generated by the OTFS wireless communication system, and the OMP algorithm is used to perform channel pre-estimation to obtain a noisy pre-estimation matrix; secondly, a hybrid void convolution and a residual path are introduced to build an improved convolutional neural network, the noisy channel estimation matrix and the actual channel matrix are input into the network, and a data-driven method is used for offline training. ； Finally, under different signal-to-noise ratios, the test data is input into the trained improved convolutional neural network, the final channel estimation matrix is ​​output, and the channel estimation performance is evaluated. The specific steps are as follows: 1) OTFS wireless communication system simulation generates data sets and preprocesses the data: First, an orthogonal amplitude modulation symbol frame of length M×N is arranged into a two-dimensional data matrix in represents an M×N-dimensional complex number set, where M and N are the number of units in the delay dimension and Doppler dimension respectively. An OTFS symbol frame is mapped to the delay-Doppler domain grid to obtain the signal x(k,l) in the delay-Doppler domain, where k∈{0,1,...,N-1}, l∈{0,1,...,M-1}; then, the time-frequency domain signal is obtained by inverse symplectic Fourier transform Where n and m represent the time domain index position and frequency domain index position of the time-frequency domain signal respectively; X[n,m] is then transformed into the time domain signal s(t) through Heisenberg transformation, which is expressed as Where t represents the sampling time of the time domain signal, g tx (t) is the time domain transmission pulse, T and Δf are the minimum sampling interval and minimum sampling frequency in the time-frequency domain respectively; then, s(t) is transmitted through the time domain channel, and the time domain signal r(t) is obtained at the receiving end, and r(t) = ∫∫h(τ,ν)g tx (t-τ)e j2 πν t dτdν+v(t), where v(t) represents additive white Gaussian noise in the time domain, h(τ,ν) represents the channel impulse response with P sparse paths, τ and v represent the path delay and Doppler frequency shift in the delayed Doppler domain, respectively. r(t) is demodulated by Wigner transform at the receiving end to obtain the time-frequency domain output signal Y[n,m]=∫g rx (t-τ)r(t)e -j2π(t-τ) dt, where g rx (t) represents the received pulse in the time domain; finally, the delayed Doppler domain output signal is obtained by sigmoid Fourier transform The delay-Doppler domain channel estimation of the OTFS communication system requires inserting a pilot symbol x into the transmitted signal. p To pre-estimate the delay Doppler domain channel, the delay Doppler domain signal after inserting the pilot can be expressed as where x d Represents data symbol, (k p ,l p ) represents the position of the pilot symbol in the delay domain and Doppler domain, k max and l max denote the maximum delay and maximum Doppler spread in the delay-Doppler domain, and p -2k max ,k p +2k max ] and l∈[l p -l max ,l p +l max ] is set to zero as the protection interval of the pilot symbol; the OMP algorithm is used for channel pre-estimation, the input pilot matrix is ​​represented as S, the delayed Doppler domain received signal is represented as Y, the starting residual r of the OMP algorithm is set to 1, the number of iterations is set to i, and the input pilot matrix S of the current iteration number i-1 is calculated i-1 All column vectors of the current residual r i-1 The product of , select the column vector s with the largest absolute value of the product i , add it to the pilot matrix of the next iteration i, denoted as S i =[S i-1 ,s i ], and the least squares method is used to obtain the channel estimation matrix of the i-th iteration The symbol ||·|| indicates taking the norm matrix, that is, Smallest As And update the residual Update the current number of iterations i = i + 1. If the relationship i>P is satisfied, stop the iteration and update the last obtained As the final estimated result Otherwise, continue to execute the above OMP algorithm operation; The dataset is generated by the OTFS wireless communication system and its size is 4×10 5 , 75% of the data is used for network training and 25% of the data is used for testing; 2) Based on the convolutional neural network, an improved convolutional neural network model is built to estimate the channel matrix The actual channel matrix H is input into the network, the normalized mean square error is used as the loss function, and the data-driven method is used for offline training; The mixed hole convolution and residual path are introduced into the convolutional neural network structure, and an improved convolutional neural network is built for noise elimination. The improved convolutional neural network includes two noise elimination modules, each of which contains three convolutional layers and a residual path. When multiple convolutional layers are superimposed in the neural network, there should be no common divisor greater than 1 between the expansion coefficients set, thereby forming a sawtooth cyclic structure. The specific training steps of the improved convolutional neural network are as follows: ① The channel pre-estimation matrix obtained by the OMP algorithm As input, the actual channel matrix H is used as the label; ② Then, three convolution layers with a convolution kernel size of 3×3 are passed through, and the expansion coefficients of these three convolution layers are set to 1, 2, and 5 respectively, forming a sawtooth cyclic structure. The improved convolutional neural network distinguishes The difference between H and realizes the noise characteristic Extraction of ③ Through the residual path and Subtract, the channel estimation result output by the output layer is ④ The above steps ②③ are regarded as the operations contained in the first noise elimination module. As the input of the second noise removal module, the actual channel matrix H is used as the label, and steps ②③ are repeated to obtain the channel estimation matrix of the improved convolutional neural network. 3) Generate test data at different signal-to-noise ratios, input the trained improved convolutional neural network, output the estimated channel matrix, and further compare it with the actual channel matrix to evaluate the channel estimation performance: After the channel estimation network training is completed, online deployment is realized. First, the OTFS communication system generates 4×10 4 The simulation data is estimated by OMP channel estimation algorithm. Then, the noise is eliminated through the channel estimation network based on the improved convolutional neural network, and finally the channel estimation matrix is ​​output. And compare it with the actual channel matrix H to evaluate the channel estimation performance.

Images