Channel estimation method based on MSM-DReEsNet in OFDM (Orthogonal Frequency Division Multiplexing) system

Abstract

The invention relates to a channel estimation method based on a multi-scale mixed depth residual error channel estimation network (MSM-DReEsNet) in an orthogonal frequency division multiplexing (OFDM) system, in particular to a channel estimation method based on a multi-scale mixed depth residual error channel estimation network (MSM-DReEsNet). According to the method, firstly, input information and output information at a pilot frequency are transmitted to a multi-scale hybrid network (MSMNet at a receiving end, the MSMNet model trains a mapping relation between the input information and the output information at the pilot frequency and channel information at the pilot frequency, then the channel information at the pilot frequency is transmitted to a deep residual error channel estimation network (DReEsNet), and the channel information at the pilot frequency is transmitted to the multi-scale hybrid network (MSMNet). The DReEsNet model trains a mapping relation from pilot frequency channel information to full channel information, and finally outputs complete channel information. Simulation results show that the provided MSM-DReEsNet model has advantages in accuracy of restoring channel state information compared with an estimation method based on ReEsNet and an estimation method based on ChannelNet.

Description

Technical Field

[0001] The present invention belongs to the technical field of channel estimation in an orthogonal frequency division multiplexing (OFDM) system, and is a channel estimation method based on a multiscale mixture deep residual channel estimation network (MSM-DReEsNet). Background Art

[0002] Orthogonal frequency division multiplexing technology is one of the most important communication systems in modern wireless communication systems. Because of the high robustness of OFDM systems to frequency selective fading and their high bandwidth utilization, OFDM has been listed as a component of the long term evolution (LTE) technology and the fifth generation new radio (5G NR) access technology standards. In wireless systems, due to the characteristics of the propagation channel, the transmitted signal is usually subject to distortion, fading and path loss. The received signal is also corrupted by additive noise and interference, further reducing the signal quality. In order to reliably recover the transmitted signal in the OFDM system, it is necessary to effectively estimate the channel information. Commonly used estimation methods include minimum mean square error (MMSE) estimation and least squares (LS) estimation. Among them, MMSE uses the second-order statistics of the channel and noise in the estimation process, which is highly complex and difficult to implement in reality. Although the estimation effect of LS estimation is not as good as that of MMSE estimation, it is widely used in engineering due to its ease of implementation.

[0003] In recent years, because image super-resolution technology is conceptually similar to the pilot-based channel estimation method in OFDM systems, convolutional neural networks (CNN)-based technology has been proposed as an alternative to traditional estimation methods. However, most neural network-based channel estimation methods perform network training on the basis of the pilot estimation value obtained by LS estimation. Since the channel information at the pilot obtained by LS estimation is not accurate enough, it affects the subsequent neural network training, resulting in inaccurate estimation of the subsequent complete channel. Summary of the invention

[0004] In view of this, the purpose of the present invention is to provide a channel estimation method based on a multi-scale mixed deep residual channel estimation network in an OFDM system. The method requires offline training of a deep residual learning network model first. The method first uses a data set generated by a wireless channel offline to train a multi-scale mixture network (MSMNet) model, and then uses the output of the MSMNet model as the input of a deep residual channel estimation network (DReEsNet) to train the DReEsNet model. In the online estimation stage, the input signal and the output signal at the pilot are first input into the MSMNet model to estimate the channel information at the pilot, and then the channel estimation information at the pilot is input into the DReEsNet model as a low-resolution noisy image for denoising and reconstruction of the channel information, and finally the received signal is equalized and demodulated to restore it to the original signal.

[0005] To achieve the above object, the present invention provides the following technical methods:

[0006] First, the MSMNet model is trained offline using the data set generated by the wireless channel. The specific structure of the model includes the following steps:

[0007] ① At the input end, the imaginary part and the real part of the input and output information of the pilot are separated to obtain a 4-dimensional real number matrix. This 4-dimensional matrix is ​​preliminarily processed through a convolution layer containing 32 convolution kernels of size 3×3 and a Relu activation function. The data is then passed to the start module (Smod) for further processing. The start module is first composed of a 3-layer structure with two convolution layers, a batch normalization (Batch Normalization, BN) layer and a Relu activation function in series, and then through a convolution layer and a batch normalization layer, a convolution layer and a batch normalization layer are also used between the input and output for jump connection to pass the shallow layer features into the deep layer to promote feature reuse. The convolution kernel size in this module is 3×3, and then it is input into the subsequent network for further feature learning;

[0008] ②After step ①, it passes through the exchange unit (EU). The exchange unit has multiple inputs and outputs. In the exchange unit, the smaller-scale feature map is converted into a larger-scale feature map through the deconvolution layer. Conversely, the larger-scale feature map is converted into a smaller-scale feature map or a feature map of the same scale through the convolution layer. Then, the feature maps of the same scale are added to obtain the corresponding scale feature output. According to the specific input and output, the specific internal structure of each exchange unit is obtained;

[0009] ③ There are three switching units in MSMNet. Each output of the first two switching units uses a residual module (Resmod). The residual module consists of two convolutional layers with 3×3 convolution kernels and a Relu activation function. The output of the last switching unit is adjusted by a convolutional layer to obtain the channel information at the pilot.

[0010] Then, the channel information at the pilot frequency estimated by the MSMNet model is used as the input of the DReEsNet model for offline training of the DReEsNet model. The structure of the DReEsNet model is as follows: the input data is first increased to 16 dimensions through a convolutional layer, then passed through 4 residual modules and a convolutional layer of the same size, and then the output of the convolutional layer is added to the output of the previous convolutional layer and upsampled to the full channel size, and finally restored to two-dimensional channel information through a convolutional layer.

[0011] In the online estimation stage, the pilot input and output information is first processed into a 4-dimensional matrix and then input into the trained MSMNet model to obtain the pilot channel estimate, whose value is Where f1(·) is the MSMNet model function, X p and Y p are the input and output signals at the pilot frequency respectively, and θ1 is the hyperparameter set of the MSMNet model.

[0012] Finally, the channel estimation value obtained at the pilot is input into the trained DReEsNet model to perform channel estimation of the complete channel, and the estimated value is Where f2(·) is the DReEsNet model function, is the channel information at the pilot estimated by the MSMNet model, and θ2 is the hyperparameter set of the DReEsNet model.

[0013] The beneficial effects of the present invention are:

[0014] The present invention relates to a channel estimation method based on MSM-DReEsNet. The method can more accurately estimate the channel information at the pilot in the OFDM channel estimation system, and can better capture the channel characteristics in the subsequent network model, so as to obtain a better channel estimation effect. Simulation results show that the method has better channel estimation performance than the channel estimation method based on the ChannelNet model in the document [1] "SOLTANI M, POURAHMADI V, MIRZAEI A, et al. Deep learning-based channel estimation [J]. IEEE Communications Letters, 2019, 23 (4): 652-655." and the channel estimation method based on the ReEsNet model in the document [2] "LI L, CHEN H, CHANG HH, et al. Deep residual learning meets OFDM channel estimation [J]. IEEE Wireless Communications Letters, 2020, 9 (5): 615-618.". Therefore, in practical applications, the method of the present invention has high utilization value and practical significance for the design of channel estimation systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to make the purpose, technical method and beneficial effects of the present invention clearer, the present invention provides the following drawings for illustration:

[0016] Figure 1 A technical roadmap for the method of the present invention;

[0017] Figure 2 This is the MSMNet model structure diagram of the method proposed in the present invention;

[0018] Figure 3 A supplementary diagram of the start module in the MSMNet model;

[0019] Figure 4 A supplementary diagram of the exchange unit in the MSMNet model;

[0020] Figure 5 This is the structural diagram of the DReEsNet model;

[0021] Figure 6 Supplementary diagram for the residual module

[0022] Figure 7 This is a comparison chart of the channel estimation performance between the MSMNet model and the LS estimation at the pilot signal;

[0023] Figure 8This is a performance comparison diagram of the method of the present invention and two other estimation methods; DETAILED DESCRIPTION

[0024] The preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0025] 1. Combine with attachment Figure 1 , Attachment Figure 2 , Attachment Figure 3 , Attachment Figure 4 , Attachment Figure 5 With attached Figure 6 It is shown that a channel estimation method based on MSM-DReEsNet in OFDM system is completed by the following steps:

[0026] Step 1: Use the data set generated by the wireless channel offline to train the MSMNet model. The specific structure of the model includes the following steps:

[0027] ① At the input end, the imaginary part and the real part of the input and output information of the pilot are separated to obtain a 4-dimensional real number matrix. This 4-dimensional matrix is ​​preliminarily processed through a convolution layer containing 32 convolution kernels of size 3×3 and a Relu activation function. Then the data is passed to the start module (start module, Smod) for further processing. The start module is first composed of 2 convolution layers, batch normalization (Batch Normalization, BN) layers and Relu activation functions. The three-layer structure is connected in series, and then through a convolution layer and a batch normalization layer, a convolution layer and a batch normalization layer are also used between the input and output for jump connection to pass the shallow layer features into the deep layer to promote feature reuse. The convolution kernel size in this module is 3×3, and then it is input into the subsequent network for further feature learning;

[0028] ②After step ①, it passes through the exchange unit (EU). The exchange unit has multiple inputs and outputs. In the exchange unit, the smaller-scale feature map is converted into a larger-scale feature map through the deconvolution layer. Conversely, the larger-scale feature map is converted into a smaller-scale feature map or a feature map of the same scale through the convolution layer. Then, the feature maps of the same scale are added to obtain the corresponding scale feature output. According to the specific input and output, the specific internal structure of each exchange unit is obtained;

[0029] ③ There are three switching units in MSMNet. Each output of the first two switching units uses a residual module (Resmod). The residual module consists of two convolutional layers with 3×3 convolution kernels and a Relu activation function. The output of the last switching unit is adjusted by a convolutional layer to obtain the channel information at the pilot.

[0030] Step 2: Use the channel information at the pilot obtained in step 1 as the input of the DReEsNet model to train the DReEsNet model offline; the DReEsNet model structure is as follows: first, the input data is increased to 16 dimensions through a convolutional layer, then passed through 4 residual modules and a convolutional layer of the same size, and then the output of the convolutional layer is added to the output of the previous convolutional layer and upsampled to the full channel size, and finally restored to two-dimensional channel information through a convolutional layer;

[0031] Step 3: In the online estimation stage, the input and output information of the pilot signal is first processed into a 4-dimensional matrix and then input into the trained MSMNet model to obtain the channel estimation of the pilot signal. The process is expressed as formula (1):

[0032]

[0033] In formula (1), f1(·) is the MSMNet model function, X p and Y p are the input and output signals at the pilot frequency, respectively, and θ1 is the hyperparameter set of the MSMNet model;

[0034] Step 4: The channel estimation value at the pilot obtained in step 3 is input into the trained DReEsNet model to perform channel estimation of the complete channel. The process is expressed as formula (2):

[0035]

[0036] In formula (2), f2(·) is the DReEsNet model function, is the channel information at the pilot estimated by the MSMNet model, and θ2 is the hyperparameter set of the DReEsNet model;

[0037] Step 5: The received information is restored based on the channel information obtained in step 3, and equalization, demodulation and other operations are performed to finally obtain the denoised information.

[0038] 2. Combine with attachment Figure 7 In order to verify that the invented OFDM channel estimation method has a better estimation performance, the model is simulated using MATLAB software. The simulation parameters are shown in Table 1:

[0039] The channel estimation performance of the MSM-DReEsNet model is compared with the channel estimation method based on the ChannelNet model in the literature [1] "SOLTANI M, POURAHMADI V, MIRZAEI A, et al. Deep learning-based channel estimation [J]. IEEE Communications Letters, 2019, 23(4): 652-655." and the channel estimation method based on the ReEsNet model in the literature [2] "LI L, CHEN H, CHANG HH, et al. Deep residual learning meets OFDM channel estimation [J]. IEEE Wireless Communications Letters, 2020, 9(5): 615-618.".

[0040] Table 1 Simulation parameter settings

[0041] Parameters Set Size Number of effective subcarriers 72 Carrier frequency 2.1GHz Channel bandwidth 1.6MHz Modulation QPSK Number of OFDM symbols 14 Number of pilots 48 Channel ETU

[0042] Under the same parameter conditions, the performance comparison of the channel estimation at the pilot point by the MSMNet model and the LS method is shown in the figure below. Figure 7 As shown in Figure 1, the MSMNet model established by the present invention has better performance than the LS pilot estimation method in the signal-to-noise ratio range of 0 to 30 dB. The minimum mean square error comparison between the signal estimated by different channel estimation models and the actual signal is shown in Figure 1. Figure 8 As shown in the figure, the MSM-DReEsNet model established in the present invention has better performance than the ChannelNet model and the ReEsNet model in the range of signal-to-noise ratio of 0 to 30 dB.

[0043] Based on the above simulation verification, it can be seen that the present invention proposes a channel estimation method based on MSM-DReEsNet in an OFDM system to address the problem that the current channel estimation of the OFDM system is not accurate enough in the channel estimation stage at the pilot, which leads to inaccurate subsequent estimation of the entire channel. The simulation results show that under the same conditions, the channel estimation scheme of the present invention can more accurately estimate the channel information at the pilot, and more accurately learn the complex mapping relationship between the pilot channel response estimation value and the entire channel response, thereby obtaining a better channel estimation effect.

[0044] Finally, it should be noted that the above preferred embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail through the above preferred embodiments, those skilled in the art should understand that various changes can be made in form and details without departing from the scope defined by the claims of the present invention.

Claims

1. A channel estimation method based on a multi-scale mixture deep residual channel estimation network (MSM-DReEsNet) in an orthogonal frequency-division multiplexing (OFDM) system; the method first uses the data set generated by the wireless channel to train the multi-scale mixture network (MSMNet) model offline, and then uses the output of the MSMNet model as the input of the deep residual channel estimation network (DReEsNet) to train the DReEsNet model. In the online estimation stage, the input signal and the output signal at the pilot are first input into the MSMNet model to estimate the channel information at the pilot, and then the channel estimation information at the pilot is used as a low-resolution noisy image and input into the DReEsNet model for denoising and reconstructing the channel information. Finally, the received signal is equalized and demodulated to restore it to the original signal. The specific steps are as follows: Step 1: Use the data set generated by the wireless channel offline to train the MSMNet model. The specific structure of the model includes the following steps: ① At the input end, the imaginary part and the real part of the input and output information of the pilot are separated to obtain a 4-dimensional real number matrix. This 4-dimensional matrix is ​​preliminarily processed through a convolution layer containing 32 convolution kernels of size 3×3 and a Relu activation function. Then the data is passed to the start module (start module, Smod) for further processing. The start module is first composed of 2 convolution layers, batch normalization (Batch Normalization, BN) layers and Relu activation functions. The three-layer structure is connected in series, and then through a convolution layer and a batch normalization layer, a convolution layer and a batch normalization layer are also used between the input and output for jump connection to pass the shallow layer features into the deep layer to promote feature reuse. The convolution kernel size in this module is 3×3, and then it is input into the subsequent network for further feature learning; ②After step ①, it passes through the exchange unit (EU). The exchange unit has multiple inputs and outputs. In the exchange unit, the smaller-scale feature map is converted into a larger-scale feature map through the deconvolution layer. Conversely, the larger-scale feature map is converted into a smaller-scale feature map or a feature map of the same scale through the convolution layer. Then, the feature maps of the same scale are added to obtain the output of the corresponding scale feature. According to the specific input and output, the specific internal structure of each exchange unit is obtained; ③ There are three switching units in MSMNet. Each output of the first two switching units uses a residual module (Resmod). The residual module consists of two convolutional layers with 3×3 convolution kernels and a Relu activation function. The output of the last switching unit is adjusted by a convolutional layer to obtain the channel information at the pilot. Step 2: Use the pilot channel information obtained in step 1 as the input of the DReEsNet model to perform offline training on the DReEsNet model; the DReEsNet model structure is as follows: first, the input data is passed through a convolutional layer to increase the dimension to 16, then passed through 4 residual modules and a convolutional layer of the same size, and then the output of the convolutional layer is added to the output of the previous convolutional layer and upsampled to the full channel size, and finally restored to two-dimensional channel information through a convolutional layer; Step 3: In the online estimation stage, the input and output information of the pilot signal is first processed into a 4-dimensional matrix and then input into the trained MSMNet model to obtain the channel estimation of the pilot signal. The process is expressed as formula (1): In formula (1), f1(·) is the MSMNet model function, X p and Y p are the input and output signals at the pilot frequency, respectively, and θ1 is the hyperparameter set of the MSMNet model; Step 4: The channel estimation value at the pilot obtained in step 3 is input into the trained DReEsNet model to perform channel estimation of the complete channel. The process is expressed as formula (2): In formula (2), f2(·) is the DReEsNet model function, is the channel information at the pilot estimated by the MSMNet model, and θ2 is the hyperparameter set of the DReEsNet model; Step 5: The received information is restored using the channel information obtained in step 3, and equalization, demodulation and other operations are performed to finally obtain the denoised information.

Images