阅读说明：本技术 一种5g高速移动系统中基于深度学习的多普勒频偏估计方法 (Doppler frequency offset estimation method based on deep learning in 5G high-speed mobile system ) 是由 杨丽花 王增浩 张捷 杨龙祥 于 2020-12-28 设计创作，主要内容包括：本发明公开了一种5G高速移动系统中基于深度学习的多普勒频偏估计方法,包括基于数据与导频信号,获取多普勒频偏初始估计,基于多普勒频偏初始估计构建预训练样本；利用构建的预训练样本对网络进行预训练处理,获取预训练网络的权重和阀值参数；基于导频信号,获取多普勒频偏初始估计,基于多普勒频偏初始估计构建训练样本；利用获取预训练网络的权重和阀值参数与构建的训练样本对网络进行第二次训练处理,获取最终的网络估计模型,进行线上多普勒频偏实时估计。具有较快的收敛速度和高的估计性能,具有一定的实用价值。(The invention discloses a Doppler frequency offset estimation method based on deep learning in a 5G high-speed mobile system, which comprises the steps of obtaining Doppler frequency offset initial estimation based on data and pilot signals, and constructing a pre-training sample based on the Doppler frequency offset initial estimation; pre-training the network by using the constructed pre-training sample to obtain the weight and threshold parameters of the pre-training network; obtaining Doppler frequency offset initial estimation based on the pilot frequency signal, and constructing a training sample based on the Doppler frequency offset initial estimation; and performing secondary training processing on the network by using the weight and threshold parameters of the obtained pre-training network and the constructed training sample to obtain a final network estimation model, and performing online Doppler frequency offset real-time estimation. The method has the advantages of high convergence rate, high estimation performance and certain practical value.)

1. A Doppler frequency offset estimation method based on deep learning in a 5G high-speed mobile system is characterized by comprising the following steps:

acquiring Doppler frequency offset initial estimation based on data and pilot frequency signals, and constructing a pre-training sample based on the Doppler frequency offset initial estimation;

pre-training the network by using the constructed pre-training sample to obtain the weight and threshold parameters of the pre-training network;

obtaining Doppler frequency offset initial estimation based on the pilot frequency signal, and constructing a training sample based on the Doppler frequency offset initial estimation;

and performing secondary training processing on the network by using the weight and threshold parameters of the obtained pre-training network and the constructed training sample to obtain a final network estimation model, and performing online Doppler frequency offset real-time estimation.

2. The method of claim 1, wherein an initial doppler frequency offset estimation is obtained based on data and pilot signals, and a formula for constructing pre-training samples based on the initial doppler frequency offset estimation is as follows:

in the formula: x_pre(U) represents the U-th input data of the BP neural network, and U is more than or equal to 0 and less than or equal to U-1; u represents the number of training sample sets; r (k) represents a received signal on the k-th subcarrier; n is a radical of_uRepresents the number of subcarriers used on one OFDMA symbol, which contains N_pA pilot frequency and N_u-N_pData; gamma-shaped_R(Z) represents a complex to real operation, indicating an initial estimate of the doppler frequency offset obtained based on the data and pilot signals.

3. The method of claim 1, wherein an initial doppler frequency offset estimation is obtained based on the pilot signal, and a formula for constructing the training samples based on the initial doppler frequency offset estimation is as follows:

in the formula: r (1) ·, R (N)_p) Is the received pilot signal.

4. The method according to any of claims 1 to 3, wherein a rice channel is used as a channel model in an HSR environment, and in a downlink transmission link, a transmission channel of an mth symbol of an ith subframe is represented as:

in the formula: c. C_iIs the LOS component of the channel;represents the scattered component and obeys a rayleigh distribution; l is the number of multipath Rice channels; epsilon_i,mIs a normalized doppler frequency offset;is the first_pNormalizing time delay of the path; n is a radical of_D＝N+N_cpIn which N is_cpIs the length of the cyclic prefix;

the Rice factor of the Rice channel is defined as

5. The Doppler frequency offset estimation method based on deep learning in a 5G high-speed mobile system according to claim 4, wherein in a 5G-NR millimeter wave HSR downlink transmission system, it is assumed that a time domain transmission signal is:

in the formula: s_i(m, k) is a signal on a k subcarrier of an mth OFDMA symbol of an ith subframe, and N is a length of fourier transform; assuming that the length of the cyclic prefix is greater than the maximum delay of the wireless transmission channel and the receiving end is in perfect timing synchronization, the received time domain signal is represented as:

in the formula:is the power of the transmission symbol; w is a_i(m, n) is a variance ofAdditive White Gaussian Noise (AWGN); because the change of the Doppler frequency offset of each OFDMA symbol on the subframe is small, the Doppler frequency offset on each subframe can be approximated to be a constant, namely epsilon_i,m≈ε_i。

6. The method of claim 1, wherein the doppler frequency offset estimation method based on deep learning in a 5G high-speed mobile system is characterized in that a pilot-assisted frequency offset estimation method is used to obtain the initial doppler frequency offset estimation.

Technical Field

The invention relates to the technical field of wireless communication, in particular to a Doppler frequency offset estimation method based on deep learning in a 5G high-speed mobile system.

Background

In recent years, rapid development of scientific technology and economy has spurred rapid and widespread development of High Speed Railways (HSRs). With the large-scale deployment of HSR, wireless communication in the HSR environment has attracted more and more attention globally, and HSR has been used as one of important usage scenarios of a fifth generation new radio (5G-NR) communication network, which is expected to achieve a mobile speed of up to 500km/h under the 5G-NR system. However, currently most wireless communication systems are designed for users with low or medium mobility, and high mobility significantly limits coverage area and transmission rate. Therefore, it is necessary to design a reliable and efficient communication system for 5G-NR HSR (>, 500km/h) scenarios.

Since the number of passengers on a high-speed train is usually large, it will be required that the HSR communication not only handle critical signaling applications, but also provide various broadband applications, such as High Definition (HD) video monitoring, on-board broadband internet access, real-time train scheduling high definition video, and railway internet of things, which require more spectrum resources to provide support. In order to meet the increasing demand of HSR user experience, some domestic and foreign documents and 3GPP standards propose that a 5G-NR millimeter wave (mmWave) frequency band is applied to HSR communication, and corresponding experimental proofs are provided. However, in a high-speed mobile environment using 5G-NR mmWave, the effect of doppler shift will become more significant and path loss more severe, which will be detrimental to vehicle-to-ground communication in an HSR scenario. Therefore, much work is required to be carried out in order to realize the 10Gbps connection of the 5G-NR millimeter waveband future train in the scene of the speed per hour up to 500 km/h.

In many of the works to be carried out, it is important to overcome the influence of large doppler shift, because in the HSR environment based on 5G-NR mmWave, the influence of doppler shift will become especially significant, such as considering the typical frequency band (30GHz) and vehicle speed 500km/h suggested by 3GPP, the doppler shift of the corresponding downlink transmission system is about 14kHz, which is 20 times that of the 4G-LTE HSR downlink system. Such a large doppler shift will cause more severe inter-carrier/user interference and cause rapid random channel changes, which severely degrades the performance of the HSR communication system. Therefore, it is necessary to develop a research on a technique for resisting large doppler shift in the HSR environment of 5G-NR mmWave, in which a doppler shift estimation and compensation technique is a base.

Disclosure of Invention

The invention aims to provide a Doppler frequency offset estimation method based on deep learning in a 5G high-speed mobile system, which has high convergence rate and high estimation performance and has certain practical value.

The invention adopts the following technical scheme for realizing the aim of the invention:

the invention provides a Doppler frequency offset estimation method based on deep learning in a 5G high-speed mobile system, which comprises the following steps:

acquiring Doppler frequency offset initial estimation based on data and pilot frequency signals, and constructing a pre-training sample based on the Doppler frequency offset initial estimation;

pre-training the network by using the constructed pre-training sample to obtain the weight and threshold parameters of the pre-training network;

obtaining Doppler frequency offset initial estimation based on the pilot frequency signal, and constructing a training sample based on the Doppler frequency offset initial estimation;

and performing secondary training processing on the network by using the weight and threshold parameters of the obtained pre-training network and the constructed training sample to obtain a final network estimation model, and performing online Doppler frequency offset real-time estimation.

Further, based on the data and the pilot signal, obtaining an initial estimation of doppler frequency offset, and constructing a formula of the pre-training sample based on the initial estimation of doppler frequency offset as follows:

in the formula: x_pre(U) represents the U-th input data of the BP neural network, and U is more than or equal to 0 and less than or equal to U-1; u represents the number of training sample sets; r (k) represents a received signal on the k-th subcarrier; n is a radical of_uRepresents the number of subcarriers used on one OFDMA symbol, which contains N_pA pilot frequency and N_u-N_pData; gamma-shaped_R(Z) represents a complex to real operation, indicating an initial estimate of the doppler frequency offset obtained based on the data and pilot signals.

Further, based on the pilot signal, obtaining an initial estimation of doppler frequency offset, and a formula for constructing a training sample based on the initial estimation of doppler frequency offset is as follows:

in the formula: r (1) ·, R (N)_p) Is the received pilot signal.

Further, a rice channel is adopted as a channel model in the HSR environment, and in a downlink transmission link, a transmission channel of an mth symbol of an ith subframe is represented as:

in the formula: c. C_iIs the LOS component of the channel;represents the scattered component and obeys a rayleigh distribution; l is the number of multipath Rice channels; epsilon_i,mIs a normalized doppler frequency offset;is the first_pNormalizing time delay of the path; n is a radical of_D＝N+N_cpIn which N is_cpIs the length of the cyclic prefix;

the Rice factor of the Rice channel is defined as

Further, in the 5G-NR millimeter wave HSR downlink transmission system, it is assumed that the time domain transmission signal is:

in the formula: s_i(m, k) is a signal on a k subcarrier of an mth OFDMA symbol of an ith subframe, and N is a length of fourier transform; assuming that the length of the cyclic prefix is greater than the maximum delay of the wireless transmission channel and the receiving end is in perfect timing synchronization, the received time domain signal is represented as:

in the formula:is the power of the transmission symbol; w is a_i(m, n) is a variance ofAdditive White Gaussian Noise (AWGN); because the change of the Doppler frequency offset of each OFDMA symbol on the subframe is small, the Doppler frequency offset on each subframe can be approximated to be a constant, namely epsilon_i,m≈ε_i。

Further, the initial estimation of the Doppler frequency offset is obtained by using a frequency offset estimation method assisted by pilot frequency.

The invention has the following beneficial effects:

offline training and online estimation with the network:

in the on-line off-training stage, the better network weight and threshold parameter are obtained through pre-training treatment, and then the network is trained again based on the obtained network parameter, so that the performance loss caused by random initialization of the network is reduced, and meanwhile, the convergence speed of the network and the prediction accuracy are improved;

in the on-line estimation stage, high-precision Doppler frequency offset estimation can be obtained in real time only by utilizing the received pilot frequency signal and Doppler frequency offset initial estimation;

has higher convergence rate and high estimation performance, thereby having certain practical value.

Drawings

FIG. 1 is a flow chart provided according to an embodiment of the present invention;

fig. 2 is a diagram of MSE performance with different numbers of training sample sets according to an embodiment of the present invention;

fig. 3 is a diagram of MSE performance obtained by training a network at different signal-to-noise ratios (SNRs) according to an embodiment of the present invention;

fig. 4 is a graph comparing MSE performance with that of the prior art at different rice factors according to an embodiment of the present invention.

Detailed Description

The invention is further described with reference to specific examples. The following examples are only for illustrating the technical solutions of the present invention more clearly, and the protection scope of the present invention is not limited thereby.

Referring to fig. 1 to 4, the technical scheme adopted by the invention is a doppler frequency offset estimation method based on deep learning applied to a 5G high-speed mobile system, and aims to improve the accuracy of doppler frequency offset estimation and reduce the computational complexity thereof. The invention firstly uses the simulation data to construct a proper training sample, then uses the training sample to carry out offline training on the BP neural network, completes the mapping relation between the input data and the output data, and finally, the Doppler frequency offset can be estimated in real time only by using the received pilot frequency signal on the online network based on the trained data. In order to reduce the performance degradation caused by random initialization, the invention designs a pre-training method to perfect the initial parameters of the BP neural network. In addition, by reasonably designing the input samples, the invention not only can utilize the characteristics in the received signals, but also can further improve the performance of Doppler frequency offset estimation by utilizing the advantages of initial estimation. The technical scheme adopted by the invention comprises the following steps:

step 1: frequency offset estimation method based on data and pilot frequency signals and assisted by pilot frequencyObtaining an initial estimate of Doppler frequency offset

Step 2: pre-training samples for building networks

In the formula, X_pre(U) U is the U input data of BP neural network, U is more than or equal to 0 and less than or equal to U-1, U is the number of training sample sets, R (k) is the received signal on the k subcarrier, N (k) is the received signal on the k subcarrier_uIs the number of subcarriers used on one OFDMA symbol, which contains N_pA pilot frequency and N_u-N_pData, Γ_R(Z) is a complex to real operation,

and step 3: pre-training the network by using the training sample constructed in the step 2;

and 4, step 4: acquiring a weight and a threshold parameter of a network convergence condition;

and 5: initial Doppler frequency offset estimation is obtained by using a pilot frequency assisted frequency offset estimation method based on pilot frequency signals

Step 6: constructing input samples for a BP neural network training phase

Wherein R is: (¹),...,R(N_p) Is a received pilot signal;

and 7: performing secondary training processing on the network by using the network weight and the threshold value obtained in the step 4 and the training sample constructed in the step 6;

and 8: acquiring a final network estimation model;

and step 9: and performing on-line Doppler frequency offset real-time estimation, and acquiring high-precision Doppler frequency offset estimation by using the received pilot frequency signal, the Doppler frequency offset estimation initial value and the trained network.

In a 5G-NR millimeter wave HSR downlink transmission system, a time domain sending signal is assumed to be

In the formula, S_i(m, k) is a signal on a k subcarrier of an mth OFDMA symbol of an ith subframe, and N is a length of fourier transform.

In the HSR communication environment, since the base stations are all built near the rails, which will have a strong direct-out (LOS) component, the rice channel is usually used as the channel model in the HSR environment, so in the downlink transmission link, the transmission channel of the mth symbol of the ith subframe can be represented as the channel model

In the formula, c_iIs the LOS component of the channel and,representing the scattered component and subject to a Rayleigh distribution, L being the number of paths of the multipath Rice channel, ε_i,mIs the normalized doppler frequency shift and is,is the first_pNormalized time delay of the path, N_D＝N+N_cpIn which N is_cpIs the length of the cyclic prefix. The Rice factor of the Rice channel is defined as

Assuming that the length of the cyclic prefix is greater than the maximum delay of the wireless transmission channel and the receiving end is in perfect timing synchronization, the received time domain signal can be expressed as

Wherein the power of the transmitted symbol isw_i(m, n) is a variance ofAdditive White Gaussian Noise (AWGN).

Because the change of the Doppler frequency offset of each OFDMA symbol on the subframe is small, the Doppler frequency offset on each subframe can be approximated to be a constant, namely epsilon_i,m≈ε_i. Since the doppler frequency offset estimation method is the same for each subframe, OFDMA symbols and subframe numbers m and i will be omitted in the present invention for convenience of introduction of the technique.

Simulation result

The performance of the invention is analyzed in conjunction with simulations. In the simulation, it is assumed that the system is an OFDMA system having one transmit antenna and one receive antenna, the OFDMA symbol length is 1024, the cyclic prefix length is 128, the carrier interval is 60kHz, the number of pilots is 32, and the distribution is concentrated. The carrier frequency is 30GHz, the vehicle speed is 500km/h, a single-path Rice channel is adopted in simulation, and Rice factors are 5 and 10. For comparison with the present invention, a doppler frequency offset estimation method based on pilot frequency segmentation and a maximum likelihood frequency offset estimation method based on pilot frequency are also provided herein.

Fig. 2 shows the MSE performance curve for different numbers of training sample sets according to the present invention. In the simulation, the number of pilots in each sample set is the same, and N_p16. It can be seen from the figure that, with the increase of the training sample set, the estimation accuracy of the doppler frequency offset estimation method obtained by the method of the present invention is higher, which indicates that a large number of training samples can improve the learning capability of the network, but will cause the increase of the complexity of offline training computation.Therefore, in practice, the number of training samples needs to be selected in a trade-off between estimation performance and computational complexity.

Fig. 3 shows the MSE performance curves obtained by training the network at different signal-to-noise ratios (SNRs) according to the present invention. When the SNR is lower than 12dB, the estimated performance obtained by training the network under the 10dB condition is better than the performance obtained by training the network under the 20dB condition. When the SNR is larger than 12dB, the estimation performance obtained by training the network under the condition of 20dB is better than that obtained by training the network at the time of 10 dB. The present invention has better estimation performance at both low and high SNRs when using a network trained at all SNR. Therefore, when performing doppler frequency offset estimation, the network can be trained according to different signal-to-noise ratios in order to obtain better performance.

Figure 4 shows the MSE performance curves for the present invention versus the prior art at different rice factors. In simulation, the invention adopts a training sample set of U-4000 and N_pWith 16 pilots, the prior art uses N_p1024 pilots and 2 pilot segments, it can be seen from the figure that the present invention can obtain better estimation performance than the prior art with fewer pilots. In addition, the present invention will have a higher performance gain as the rice factor increases.

The above description is only a preferred embodiment of the present invention, and it should be noted that, for those skilled in the art, several modifications and variations can be made without departing from the technical principle of the present invention, and these modifications and variations should also be regarded as the protection scope of the present invention.