5G NR frequency fading channel estimation method and system based on deep learning assistance

Abstract

The invention relates to the technical field of wireless communication, and particularly discloses a 5G NR frequency fading channel estimation method and system based on deep learning assistance, a channel estimation model based on deep learning assistance is constructed, the model adopts a low-level feature extraction network to learn low-level features of an LS channel response matrix, and the low-level features of the LS channel response matrix are extracted. The method comprises the following steps of: firstly, extracting global correlation features of a time domain and a frequency domain by using a multi-scale feature extraction network, so that the state perception of a 5G NR receiver channel is more comprehensive, finally, improving the fitting capability in combination with an output layer to solve the problem of insufficient noise estimation of a traditional LS algorithm, and improving the accuracy of the state information of a receiver communication channel at a high frequency band and different delays. Simulation experiments show that the channel response mean square error is obviously superior to the LS algorithm, and the robustness is higher than that of a linear minimum mean square error algorithm.

Description

Technical Field

[0001] The present invention relates to the field of wireless communication technologies, and in particular, to a 5G NR frequency fading channel estimation method and system assisted by deep learning. Background Art

[0002] In the face of the intelligent, high-speed transmission, and diverse service requirements of 5G, the accuracy of the channel state information (CSI) of wireless communication is particularly important. In an actual communication environment, the 5G channel is prone to non-stationary losses due to multipath transmission, environmental noise, etc. Moreover, in the 5G NR (New Radio) system, the channel bandwidth is large, the amount of data carried by subcarriers is large, the dimension of the channel estimation data matrix is high, and the anti-interference ability is weak. Demodulation reference signals (DMRS) are inserted into the subcarriers of orthogonal frequency division multiplexing (OFDM) symbols to obtain accurate information. However, in order to save spectrum resources, it is only transmitted on some subcarriers, which adds great difficulty to channel estimation. Most of the channel estimation methods used in 5G NR are the least square (LS) algorithms, and the estimation time is fast. The LS algorithm of 5G NR uses two-dimensional interpolation of channel information at pilot positions for channel estimation, but ignores channel noise, so the performance is poor.

[0003] Due to the rapid development of software and hardware, deep learning technology has been maturely applied in many fields, and the application of deep learning in wireless communication is also very popular, moving towards the intelligent direction. Some literature has proposed a data-driven channel estimation method based on a super-resolution network, regarding the least square estimated channel as an image polluted by noise, and using the channel sparsity and the attention mechanism of the super-resolution network to effectively recover channel information, which has better performance than traditional algorithms. However, the processed data is a single OFDM symbol, and 5G NR sends data in frames, and there are a large number of pilots in each frame. The pilots of a single OFDM symbol cannot verify the channel estimation of each frame.

[0004] Given that the structural feature correlation of the channel response matrix in time-frequency points is large and the deep learning technology has strong feature extraction ability, more attention is paid to noise processing. Some literature proposes to model the channel matrix in a large-scale multiple-input multiple-output (MIMO) system as a 2D image and use a denoising network based on a convolutional neural network (CNN) for channel estimation. Some literature configures an OFDM system with reference to the 3GPP protocol and proposes a deep convolutional estimation network based on the MLP-mixer structure. Using the estimation of the LS algorithm as the initial feature matrix, it learns to recover the LS performance of the features. The simulation results show that its algorithm is superior to the traditional algorithm. Some literature proposes an image restoration network based on the combination of LS-based deep convolution neural network and residual connection for the 5G NR OFDM system. The simulation results show that the performance is superior to the traditional algorithm in a time-varying channel. Although the data in these literatures refer to the 5G NR frame structure and are relatively real, they are all superior in performance in a time-varying channel, but the LS estimation recovery performance is poor in a frequency-selective fading channel, and the model training effect is not good. Summary of the Invention

[0005] The present invention provides a method and system for 5G NR frequency fading channel estimation assisted by deep learning, and the technical problem to be solved is that: the existing technology has insufficient estimation of the LS algorithm in a frequency-selective fading channel caused by multipath loss under the high-rate requirements of 5G NR.

[0006] To solve the above technical problems, the present invention provides a method for 5G NR frequency fading channel estimation assisted by deep learning, including the steps of: Obtain the LS channel response matrix at the receiving end of the 5G NR system; Input the LS channel response matrix into a channel estimation model assisted by deep learning, and the channel estimation model outputs a predicted channel response matrix.

[0007] Further, the channel estimation model includes a low-level feature extraction network, a multi-scale feature extraction network, and an output layer. The LS channel response matrix is input into the low-level feature extraction network, and after matrix processing and local feature acquisition, multiple layers of low-level features are obtained and input into the multi-scale feature extraction network; the multi-scale feature extraction network performs dilated convolution on the multi-level low-level features and then performs multi-level feature fusion to obtain one-dimensional multi-scale features, and outputs a predicted channel response matrix through the output layer.

[0008] Further, the processing flow of the low-level feature extraction network includes the steps of: For the LS channel response matrix Perform channel matrix processing to obtain single-channel data; Perform multi-layer convolution on the single-channel data to obtain multi-layer convolution outputs; Independently perform a normalization operation on the multi-layer convolution outputs and then perform a max pooling operation, and finally perform an activation process and a global average pooling to obtain corresponding multi-layer low-level features.

[0009] Furthermore, the processing flow of the multi-scale feature extraction network includes the steps of: Perform parallel convolution on the multi-layer low-level features according to the levels to obtain multi-dimensional convolution features, where at least two layers are dilated convolutions with different dilation rates; Concatenate the extracted multi-dimensional convolution features to obtain a one-dimensional multi-scale feature.

[0010] Furthermore, the low-level feature extraction network performs four-layer convolution on the single-channel data to obtain four-layer convolution outputs, where the output of the equals , is the convolution operation function, is the convolution kernel weight matrix of the layer, and valid specifies that the type of convolution operation is valid, is the bias parameter.

[0011] Furthermore, the low-level feature extraction network performs a normalization operation on the output of the layer convolution, and the obtained normalization result equals , represents the mean of the output of the layer convolution , is the variance of the output of the layer convolution , represents a constant to prevent the denominator from being zero, , respectively represent weights and biases.

[0012] Furthermore, the multi-scale feature extraction network sets four convolutional layers to perform parallel convolution, where the first layer is a 1×1 convolution, and the second to fourth layers are three dilated convolutions with different dilation rates of convolution, .

[0013] Furthermore, during the training process, the mean square error between the predicted output and the true output is used as the loss.

[0014] The present invention also provides a 5G NR frequency fading channel estimation system based on deep learning assistance, which is characterized in that the system includes an LS channel response module and a channel estimation module; The LS channel response module is used to obtain the LS channel response matrix at the receiving end of the 5G NR system; The channel estimation module is used to input the LS channel response matrix into a deep learning-assisted channel estimation model, and the channel estimation model outputs a predicted channel response matrix.

[0015] Preferably, the channel estimation model includes a low-level feature extraction network, a multi-scale feature extraction network, and an output layer; the low-level feature extraction network is used to obtain multi-layer low-level features after matrix processing and local feature acquisition of the LS channel response matrix, and input them into the multi-scale feature extraction network; the multi-scale feature extraction network is used to perform dilated convolution on the multi-level low-level features and then perform multi-level feature fusion to obtain one-dimensional multi-scale features; the output layer is used to output a predicted channel response matrix according to the one-dimensional multi-scale features ; The low-level feature extraction network includes a matrix processing module, a convolutional layer, a normalization layer, a pooling layer, and a non-linear layer; the matrix processing module is used to perform channel matrix processing on the LS channel response matrix to obtain single-channel data; the convolutional layer is used to perform multi-layer convolution on the single-channel data to obtain multi-layer convolution outputs; the normalization layer and the pooling layer are used to independently perform a normalization operation on the multi-layer convolution outputs and then perform a max-pooling operation, and the non-linear layer finally performs activation processing and global average pooling to obtain corresponding multi-layer low-level features; The multi-scale feature extraction network includes a parallel dilated convolutional layer and a multi-level feature fusion module; the parallel dilated convolutional layer is used to perform parallel convolution on the multi-layer low-level features according to levels to obtain multi-dimensional convolutional features, where at least two layers are dilated convolutions with different dilation rates; the multi-level feature fusion module is used to splice the extracted multi-dimensional convolutional features to obtain one-dimensional multi-scale features.

[0016] The 5G NR frequency fading channel estimation method and system based on deep learning assistance provided by the present invention constructs a channel estimation model based on deep learning assistance. This model uses a low-level feature extraction network to learn the low-level features of the LS channel response matrix, and then uses a multi-scale feature extraction network to learn the features of the global correlation in the time domain and frequency domain, making the state perception of the 5G NR receiver channel more comprehensive. Finally, the fitting ability is improved by combining the output layer to solve the problem of insufficient noise estimation in the traditional LS algorithm, and the accuracy of the communication channel state information of the receiver in the high frequency band and different delays is improved. Simulation experiments show that the mean square error (MSE) of the channel response of the present invention is significantly better than that of the LS algorithm, and it has higher robustness than the linear minimum mean square error (LMMSE) algorithm. Description of the Drawings

[0017] Figure 1 It is a conceptual diagram of a resource grid RB and a RE provided by an embodiment of the present invention; Figure 2 It is a structural overview diagram of a channel estimation system model provided by an embodiment of the present invention; Figure 3 It is a flowchart of a low-level feature extraction network provided by an embodiment of the present invention; Figure 4 It is a refined structural diagram of a channel estimation system model provided by an embodiment of the present invention; Figure 5 It is an example diagram of the actual receptive field of a dilated convolution kernel provided by an embodiment of the present invention; Figure 6 It is a flowchart of the simulation model modeling of the 5G NR receiving end provided by an embodiment of the present invention; Figure 7 It is an OFDM modulation waveform diagram provided by an embodiment of the present invention; Figure 8 It is a training and verification MSE loss diagram provided by an embodiment of the present invention; Figure 9 It is an amplitude and phase diagram of the first frame LS and the true channel response provided by an embodiment of the present invention; Figure 10 It is a channel estimation performance comparison diagram of three algorithms when the signal-to-noise ratio is 5 dB provided by an embodiment of the present invention; Figure 11 It is a channel estimation performance comparison diagram of three algorithms when the signal-to-noise ratio is 10 dB provided by an embodiment of the present invention; Figure 12 It is a channel estimation performance comparison diagram of three algorithms when the signal-to-noise ratio is 15 dB provided by an embodiment of the present invention; Figure 13It is a comparison chart of the channel estimation performance of three algorithms when the signal-to-noise ratio is 20 dB provided by an embodiment of the present invention; Figure 14 It is a comparison chart of the channel estimation performance of three algorithms when the signal-to-noise ratio is 25 dB provided by an embodiment of the present invention. Detailed implementation manners

[0018] The following specifically clarifies the implementation manners of the present invention in conjunction with the accompanying drawings. The given embodiments are only for illustrative purposes and should not be construed as a limitation of the present invention. The accompanying drawings are only for reference and illustration and do not constitute a limitation on the protection scope of the present invention, because many changes can be made to the present invention without departing from the spirit and scope of the present invention. Embodiment

[0019] An embodiment of the present invention provides a 5G NR frequency fading channel estimation method assisted by deep learning, including the steps: S1. Obtain the LS channel response matrix at the receiving end of the 5G NR system; S2. Input the LS channel response matrix into a channel estimation model assisted by deep learning, and the channel estimation model outputs a predicted channel response matrix.

[0020] The present invention builds a 5G NR communication system for data collection, uses the channel estimation matrix of the system as a feature set that needs to process noise, regards the actual channel matrix as a result containing noise, uses the matrix obtained by LS two-dimensional interpolation as the input of the problem, and the real channel as the output, and models the LS channel estimation problem as a problem of fitting the input and output. Among them, the channel estimation model combines a convolutional neural network (CNN) and an atrous spatial pyramid pooling layer (ASPP). The convolutional neural network performs low-level feature learning, and the ASPP module learns the features of global correlation in the time domain and frequency domain. Finally, a fully connected layer is combined to improve the fitting ability to recover the noise ignored by the traditional algorithm LS.

[0021] The implementation scenario of this embodiment is a 5G NR communication scenario. Frame signals are generated according to the 5G NR standard protocol, and the physical downlink shared channel (PDSCH) transmission signal of single-input single-output (SISO) is considered. The receiving end of 5G NR obtains a complex frequency domain matrix through the high-order modulation mapping operation of the transmitter , the m th complex frequency domain ( is in the complex number domain) is a complex number, m = 1, 2,..., M。Moreover, through the 5G NR channel with the non-stationary characteristics of frequency-selective fading, the interference of each symbol increases, and the LS channel estimation algorithm has poor performance.

[0022] At the 5G NR transmitter, data is transmitted through Resource Elements (REs), and each RE corresponds to a subcarrier on an OFDM symbol; according to the NR standard protocol, one time slot consists of (14 in this example) OFDM symbols, and each time slot is sent as a subframe Resource Blocks (RBs), where each resource block contains (i.e., 12) subcarriers. The dimension of the signal matrix transmitted in this example is defined by the resource grid concept diagram shown in Figure 1 .

[0023] Figure 1 One RB in has a total of REs. The blue RE grid represents the received signal information, and the green RE represents the Demodulation Deference Signals (DMRS), that is, the pilot signal. Thus, the first dimension of the generated signal matrix is the number of subcarriers, the second dimension is the number of OFDM symbols in the time domain, and the third dimension is the number of transmitted. Then, the resource information is obtained by the receiving antenna, the frame signal is parsed, and the data is matrix-transformed to convert the complex data into two paths, one for the real part and one for the imaginary part, and then the allocated RB blocks are taken out in sequence and spliced into a long sequence.

[0024] Based on this data matrix, the signal data at the DMRS RE during transmission is extracted and the LS algorithm is used to utilize this reference information, and finally the channel response at the pilot is obtained: (1), where, represents the difference of the object to be obtained, represents the estimated pilot response, and represent the received pilot information and the transmitted DMRS value, represents the channel estimation value of the symbol at the pilot in the resource grid.

[0025] The minimum error between the actual observed value and the estimated observed value is obtained to get the channel response matrix at the DMRS position: (2).

[0026] Finally, the following two-dimensional linear interpolation algorithm in Equation (3-5) can be used to obtain the entire channel response .

[0027] (3), (4), (5), In formulas (3 - 5), represents a two - dimensional interpolation function, represents the grid matrix generated by the position indices of the pilot information in the time domain and frequency domain, and respectively represent obtaining the channel response amplitude and phase of the pilot position, represents the grid formed by the time domain and frequency domain of the resource grid, represents that the interpolation type is linear interpolation, represents the entire channel frequency response finally obtained by two - dimensional interpolation using the LS algorithm.

[0028] The channel response ignores the influence of noise, resulting in inaccurate channel state prediction. Therefore, a channel estimation model assisted by deep learning is adopted to learn the noise characteristics not considered. In the model, a convolutional neural network is used to extract data - related features and a parallel convolutional layer is constructed to further obtain multi - scale global information, and finally the predicted channel response is fused and output.

[0029] The objective function for training the channel estimation model is shown as the following formula: (6), This objective function represents the convergence of and the predicted channel response the minimum value of the error between them to obtain excellent channel state information under 5G NR frequency - selective fading. Therefore, the parameters of the LS - based feature - fusion channel estimation model are analyzed in detail below and the network structure is reasonably designed.

[0030] Because the time - slot data at the receiving end transmits more than just a single RB block of data, and the pilot data in the data occupies 1 / 10 of the data as shown in Figure 1 . Therefore, the channel matrix data is large and the feature relationship is complex, making it difficult for traditional neural networks to obtain good channel estimation results. Therefore, it is necessary to ensure multi - scale and multi - level feature extraction. Therefore, the channel estimation model designed in the embodiments of the present invention includes three parts: a low - level feature extraction network, a multi - scale feature extraction network, and an output layer, as shown in Figure 2 . The LS channel response matrix The input low-level feature extraction network (Convolutional Neural Network, CNN) obtains low-level features after matrix processing and local feature acquisition, and then inputs the low-level features into the multi-scale feature extraction network (ASPP). The multi-scale feature extraction network performs dilated convolution on the low-level features and then conducts multi-level feature fusion, and finally outputs the predicted channel response matrix through the output layer. .

[0031] The feature extraction process of the low-level feature extraction network is as Figure 3 shown, and specifically includes the steps: S21. Process the LS channel response matrix to obtain single-channel data; S22. Perform multi-layer convolution on the single-channel data to obtain multi-layer convolution output; S23. Independently perform a normalization operation on the multi-layer convolution output, then perform a max-pooling operation, and finally perform an activation process and a global average pooling to obtain the corresponding multi-layer low-level features.

[0032] In step S21, according to the system model, the data of the channel response matrix is in complex form, and the current CNN neural network architecture cannot directly process complex numbers. Therefore, matrix processing is required to extract the real and imaginary parts of the complex numbers. According to their strong data correlation, a single-channel data is formed and then input into the CNN network model for feature learning.

[0033] Steps S22 and S23 are implemented using the Convolutional Neural Network (CNN). The intermediate structure layer of the CNN neural network is the main part of the low-level feature extraction of the 5GNR channel. The convolution of each layer of the feature extraction unit is shown in the following formula (7): (7), where, is the output of the th layer after this convolution operation, is the convolution operation function, is the convolution kernel weight matrix of the th layer, indicates that the type of convolution operation is (no padding is performed on the input data during the convolution process, and the convolution kernel only performs convolution operations when it is completely inside the input data), is the bias parameter.

[0034] After convolution, a non-linear transformation is performed through the batch normalization layer as shown in the following formula (8): (8), where, represents the Mean of the convolutional output of the layer, is the variance of the convolutional output of the layer, represents a small constant to prevent division by zero, , represent the weight and bias respectively. represents the output of the layer after normalization.

[0035] Then it passes through the max - pooling layer with the kernel size set to , stride of 2, and the dimension of the output channel matrix is halved. Finally, the ReLu activation function is used for non - linear processing to prevent gradient vanishing or gradient explosion.

[0036] As Figure 2 and Figure 4 shown in the left - hand part, the specific structure of the low - level feature extraction network is: successively cascading a convolutional layer, a normalization layer, a pooling layer, and an activation layer. For each feature unit, the output of the th unit first passes through a convolutional layer with the number of convolutional filters being . At this time, the size of the convolutional kernel used is 3×3, the padding value is 1, the output feature dimension remains unchanged. After normalization as in Equation (8), max - pooling is performed to focus on the most prominent positions of the features, and the dimension of the output matrix is halved. Then, in the activation layer, the ReLU function is used for non - linear processing to accelerate the convergence speed of the model. Finally, through global average pooling, the low - level features of the 5G NR channel are obtained.

[0037] Because the data dimension of the channel matrix is two - dimensional and has original features such as a multi - scale feature dataset, features can be extracted by the dilated convolutional kernel shown in Figure 5 . The multi - scale feature extraction network ASPP uses parallel dilated convolutions for multi - level associated information analysis to obtain features with different receptive field sizes, and fully extracts global feature information and local feature information. The dilated convolutional kernel in the multi - scale feature extraction network ASPP has a parameter - dilation rate (rate), whose meaning is that on the basis of ordinary convolution, the interval between adjacent weights is set to rate - 1. The default rate of ordinary convolution is 1, which represents a receptive field with holes.

[0038] As Figure 2 and Figure 4 shown in the right - hand part, the specific operation steps of the multi - scale feature extraction network ASPP adopted in this embodiment are as follows: S24. Perform parallel convolution on the multi - layer low - level features according to the levels to obtain multi - dimensional convolutional features, where at least two layers are dilated convolutions with different dilation rates.

[0039] In this embodiment, the low-level feature extraction network has four convolutional layers. Correspondingly, the multi-scale feature extraction network sets four convolutional layers for parallel convolution, where the first layer is a 1×1 convolution, and the second to fourth layers are three convolutional layers with different dilation rates of convolution. To adapt to the dimensions of the convolutional output matrix, this example sets , , which can better obtain global information. The actual kernel size is: (9), where represents the size of the actual dilated convolutional kernel, and the actual convolution size can be obtained through this formula.

[0040] Performing parallel convolution operations on the original low-level features, the operation process is the same as that of formula (7), and the size of the convolutional output matrix is obtained as follows: (10), where and are the sizes in the horizontal and vertical axes respectively, and represent the width and height of the input feature matrix respectively, , represent the padding in the horizontal and vertical axes respectively, , represent the strides in the horizontal and vertical axes respectively.

[0041] S25. Extracting multi-dimensional convolutional features through the splicing of a convolutional kernel with a kernel size of , placing the dimensions into one dimension to obtain one-dimensional multi-scale features. The obtained multi-scale features can represent the global correlation information of the input low-level features.

[0042] Finally, a fully connected layer is used to perform matrix output on the one-dimensional multi-scale features to obtain the prediction matrix .

[0043] The poor channel response effect is caused by the incomplete acquisition of data feature information by the CNN. Through the ASPP, the state perception of the 5G NR receiver channel is more comprehensive. Without significantly increasing the computational complexity, a channel estimation model composed of the CNN and the ASPP is constructed to improve the accuracy of the communication channel state information of the high-frequency band and receivers with different delays, and further optimize the function of the 5G NR receiver channel estimation module. The entire process of the channel estimation model can be expressed by the formula: (11), where , respectively represent the training parameters of CNN and ASPP, represents the functional expression of CNN, represents the functional expression of ASPP.

[0044] To define the error between the predicted output and the true output, this model uses the mean square error (MSE) of the entire training sample as the loss, which is expressed as follows: (12), where, refers to the number of training samples in each training round of the training set, respectively refer to the th estimated and actual channel frequency responses in this round.

[0045] It should be noted that various forms of the process shown above can be used, reordering, adding or deleting steps. For example, the steps described in the present invention can be executed in parallel, sequentially or in different orders, as long as the desired results of the technical solution of the present invention can be achieved. This embodiment does not limit this here.

[0046] Corresponding to the above-mentioned 5G NR frequency fading channel estimation method assisted by deep learning, this embodiment also provides a 5G NR frequency fading channel estimation system assisted by deep learning. Referring to Figures 2 to 5 , this system includes an LS channel response module and a channel estimation module; The LS channel response module is used to obtain the LS channel response matrix at the receiving end of the 5G NR system; The channel estimation module is used to input the LS channel response matrix into the channel estimation model assisted by deep learning, and the channel estimation model outputs the predicted channel response matrix.

[0047] Among them, the channel estimation model includes a low-level feature extraction network, a multi-scale feature extraction network and an output layer; the low-level feature extraction network is used to obtain multi-layer low-level features after matrix processing and local feature acquisition of the LS channel response matrix and input them into the multi-scale feature extraction network; the multi-scale feature extraction network is used to perform dilated convolution on the multi-level low-level features and then perform multi-level feature fusion to obtain one-dimensional multi-scale features; the output layer is used to output the predicted channel response matrix according to the one-dimensional multi-scale features ; The low-level feature extraction network includes a matrix processing module, a convolutional layer, a normalization layer, a pooling layer and a non-linear layer; the matrix processing module is used to Perform channel matrix processing to obtain single-channel data; the convolutional layer is used to perform multi-layer convolution on the single-channel data to obtain multi-layer convolutional outputs; the normalization layer and the pooling layer are used to independently perform normalization operations on the multi-layer convolutional outputs and then perform max pooling operations, and the non-linear layer finally performs activation processing and global average pooling to obtain corresponding multi-layer low-level features; The multi-scale feature extraction network includes a parallel dilated convolutional layer and a multi-level feature fusion module; the parallel dilated convolutional layer is used to perform parallel convolution on the multi-layer low-level features according to levels to obtain multi-dimensional convolutional features, where at least two layers are dilated convolutions with different dilation rates; the multi-level feature fusion module is used to splice the extracted multi-dimensional convolutional features to obtain one-dimensional multi-scale features.

[0048] The specific operations of each module and network layer of the system have been described in the above method and will not be elaborated in this system.

[0049] It should be emphasized that the embodiments described in the present invention can be implemented in a computing system including backend components (e.g., as a data server), or a computing system including middleware components (e.g., an application server), or a computing system including frontend components (e.g., a user computer having a graphical user interface or a web browser through which the user can interact with the embodiments of the systems and technologies described herein), or a computing system including any combination of such backend components, middleware components, or frontend components. The components of the system can be interconnected with each other through digital data communication in any form or medium (e.g., a communication network). Examples of communication networks include: local area network (LAN), wide area network (WAN), blockchain network, and the Internet.

[0050] The computer programs for implementing the methods and systems of the present invention can be written in any combination of one or more programming languages and stored in a computer-readable storage medium. These computer programs can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing devices, such that when the computer programs are executed by the processor, the functions / operations specified in the flowcharts and / or block diagrams are implemented. The computer programs can be executed entirely on the machine, partially on the machine, as an independent software package partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0051] A computer-readable storage medium can be a tangible medium that can contain or store a computer program for use by or in connection with an instruction execution system, apparatus, or device. The computer-readable storage medium can include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. Alternatively, the computer-readable storage medium can be a machine-readable signal medium. More specific examples of the machine-readable storage medium would include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, compact disc read-only memory (CD ROM), optical storage devices, magnetic storage devices, or any combination of the foregoing.

[0052] In summary, the 5G NR frequency fading channel estimation method and system provided by the present invention constructs a channel estimation model assisted by deep learning. This model uses a low-level feature extraction network to learn the low-level features of the LS channel response matrix, and then utilizes a multi-scale feature extraction network to learn the features of the global correlation in the time domain and frequency domain, making the state perception of the 5G NR receiver channel more comprehensive. Finally, by combining the output layer to enhance the fitting ability, the problem of insufficient noise estimation in the traditional LS algorithm is solved, and the accuracy of the communication channel state information of the receiver in the high frequency band and with different delays is improved.

[0053] The following is the simulation and result analysis.

[0054] In this simulation, subcarriers are generated according to the 5G NR standard, and the system modeling process is as Figure 6 shown. A PDSCH simulation model is built using the Pytorch 2.3.1 framework and the Python 3.11.9 program. The PDSCH generates a resource grid to transmit data between the user equipment (UE) and the base station (BS). Set the number of subcarriers of a resource grid RB of the system to 12, the subcarrier spacing to 15 kHz, the carrier frequency to 2.4 GHz, the number of carriers per RB = 12, and the occupied bandwidth to 180 kHz. The demodulation reference signal DMRS is set at the 2nd, 5th, 8th, and 11th indices of the OFDM symbol and every 3rd subcarrier. According to the protocol, one radio frame is 10 subframes, each subframe has one time slot, each time slot has 14 OFDM symbols, and each time slot is limited to 106 RBs, which can be flexibly set. Set to 84. In this paper, the size of the channel matrix of one time slot is a complex number of 1008×14.

[0055] The resource grid generated at the transmitting end undergoes 64QAM modulation, layer mapping, and then resource mapping, followed by OFDM modulation using the Inverse Fast Fourier Transform (IFFT). According to the protocol, the FFT point number is set to 1024 points. In modulation, a cyclic prefix (CP) with a length of 144 is added to avoid multipath interference between symbols. The waveform signal generated after modulation is as shown in Figure 7 shown. To meet the frequency attenuation characteristics, a Rayleigh fading channel with a complex Gaussian distribution is considered in this paper, with a standard deviation set to 0.5, and 5 values conforming to the Rayleigh distribution are randomly generated. After passing through the channel, the signal amplitude will be attenuated and noise will be generated. In this paper, additive white Gaussian noise is added, and after removing the CP and performing FFT time-frequency demodulation, channel estimation can be carried out.

[0056] A neural network model is built using the Pytorch framework, and three datasets, namely the training set, validation set, and test set, are set. To prevent overfitting, the early stopping method is adopted, and the patience is set. When it is exceeded, the best model is saved. The parameter settings of other training models are shown in Table 1. The low-level CNN convolutional layer of the model in this paper is set to 4, and the ASPP high-level features are set to be parallel with four, one global feature convolutional layer, and three dilated convolutional layers.

[0057]

[0058] During the network training and validation process, to better fit the influence of noise on LS channel estimation, data with a signal-to-noise ratio of 25dB is used for model training. The losses during training and validation are as shown in Figure 8 shown. The best model appears in the 73rd round, and this best model is saved for online testing.

[0059] Under the same conditions, the LS channel estimation matrix with a signal-to-noise ratio of 25dB and 100 frames is sent online, along with the true channel frequency response of the current channel. The LS estimation matrix is used as the input of the CNAP algorithm model, and the true channel response is used as the output. Since the two-dimensional interpolation used by LS is interpolated from the amplitude and phase dimensions and a total of 100 frames are sent, the data volume is too large, so the corresponding amplitude and phase of the first frame are given as shown in Figure 9 shown.

[0060] Figure 10 、 Figure 11 、 Figure 12 、 Figure 13 、 Figure 14Shows the comparison of the minimum mean square error (MSE) channel estimation performance of the channel estimation algorithm (CNAP) proposed in this embodiment with the traditional LS algorithm and the traditional LMMSE algorithm when the signal-to-noise ratios are 5 dB, 10 dB, 15 dB, 20 dB, and 25 dB, taking the average every 10 frames for 100 frames of data. Comparison Figures 10 to 14 It can be seen that: for the LS algorithm, as the signal-to-noise ratio increases, its minimum mean square error decreases, but the error is still too large, and the mean square error fluctuates too much up and down; while the LMMSE algorithm performs linear smoothing of noise, so its performance is greatly improved compared with the LS algorithm, but its mean square error also fluctuates relatively large; the CNAP algorithm proposed in this embodiment is significantly better than the LS algorithm in terms of performance as the signal-to-noise ratio gradually increases, and as the signal-to-noise ratio increases, its performance also changes from being comparable to the LMMSE algorithm to being better than the LMMSE algorithm. Moreover, the CNAP algorithm proposed in this embodiment has the flattest fluctuation trend and the most stable performance among the mean square errors at all signal-to-noise ratios.

[0061] The simulation experiment shows that the 5G NR channel estimation performance is better than the LS and LMMSE algorithms at different signal-to-noise ratios, and the performance is very stable and has higher robustness.

[0062] The above embodiments are the preferred embodiments of the present invention, but the embodiments of the present invention are not limited by the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications made without departing from the spirit and principle of the present invention shall be equivalent replacement methods and are all included in the protection scope of the present invention.

Claims

1. A 5G NR frequency fading channel estimation method assisted by deep learning, characterized in that Including the steps: Obtain the LS channel response matrix of the receiving end of the 5G NR system; Input the LS channel response matrix into a deep learning-assisted channel estimation model, and the channel estimation model outputs a predicted channel response matrix.

2. The 5G NR frequency fading channel estimation method assisted by deep learning according to claim 1, characterized in that: The channel estimation model includes a low-level feature extraction network, a multi-scale feature extraction network, and an output layer. The LS channel response matrix is input into the low-level feature extraction network, and after matrix processing and local feature acquisition, multi-layer low-level features are obtained and input into the multi-scale feature extraction network; the multi-scale feature extraction network performs dilated convolution on the multi-level low-level features and then conducts multi-level feature fusion to obtain one-dimensional multi-scale features, which are output as the predicted channel response matrix via the output layer.

3. The 5G NR frequency fading channel estimation method assisted by deep learning according to claim 2, characterized in that, The processing flow of the low-level feature extraction network includes the steps: For the LS channel response matrix Perform channel matrix processing to obtain single-channel data; Perform multi-layer convolution on the single-channel data to obtain multi-layer convolution outputs; Independently perform a normalization operation on the multi-layer convolution outputs, then perform a max-pooling operation, and finally perform an activation process and a global average pooling to obtain the corresponding multi-layer low-level features.

4. The 5G NR frequency fading channel estimation method assisted by deep learning according to claim 3, characterized in that The processing flow of the multi-scale feature extraction network includes the steps: Perform parallel convolution on the multi-layer low-level features by level to obtain multi-dimensional convolution features, where at least two layers are dilated convolutions with different dilation rates; Concatenate the extracted multi-dimensional convolution features to obtain one-dimensional multi-scale features.

5. The 5G NR frequency fading channel estimation method assisted by deep learning according to claim 4, wherein: The low-level feature extraction network performs four-layer convolution on the single-channel data to obtain four-layer convolution outputs, where the equals , is the convolution operation function, is the convolution kernel weight matrix of the th layer, and valid indicates that the type of convolution operation is valid, is the bias parameter.

6. The 5G NR frequency fading channel estimation method assisted by deep learning according to claim 5, wherein: The low-level feature extraction network performs a normalization operation on the convolution output of the th layer, and the obtained normalization result is equal to , denotes the mean of the convolution output of the th layer, is the variance of the convolution output of the th layer, , denotes a constant to prevent the denominator from being zero, , denote the weight and bias respectively.​ 7. The 5G NR frequency fading channel estimation method assisted by deep learning according to claim 6, characterized in that: The multi-scale feature extraction network is set with four convolutional layers for parallel convolution, where the first layer is a 1×1 convolution, and the second to fourth layers are convolutions with three different dilation rates convolutions, .

8. The 5G NR frequency fading channel estimation method assisted by deep learning according to claim 7, characterized in that: During the training process, use the mean square error between the predicted output and the true output as the loss.

9. A 5G NR frequency fading channel estimation system assisted by deep learning, characterized in that, Including an LS channel response module and a channel estimation module; The LS channel response module is used to obtain the LS channel response matrix of the receiving end of the 5G NR system; The channel estimation module is used to input the LS channel response matrix into a deep learning-assisted channel estimation model, and the channel estimation model outputs a predicted channel response matrix.

10. The 5G NR frequency fading channel estimation system assisted by deep learning according to claim 9, wherein: The channel estimation model includes a low-level feature extraction network, a multi-scale feature extraction network, and an output layer; the low-level feature extraction network is used to perform matrix processing and local feature acquisition on the LS channel response matrix to obtain multiple layers of low-level features and input them into the multi-scale feature extraction network; the multi-scale feature extraction network is used to perform dilated convolution on the multi-level low-level features and then perform multi-level feature fusion to obtain one-dimensional multi-scale features. The output layer is used to output a predicted channel response matrix according to the one-dimensional multi-scale features ; The low-level feature extraction network includes a matrix processing module, a convolutional layer, a normalization layer, a pooling layer, and a non-linear layer; the matrix processing module is used to perform channel matrix processing on the LS channel response matrix to obtain single-channel data; the convolutional layer is used to perform multi-layer convolution on the single-channel data to obtain a multi-layer convolution output; the normalization layer and the pooling layer are used to independently perform a normalization operation on the multi-layer convolution output and then perform a max pooling operation, and the non-linear layer finally performs an activation process and a global average pooling to obtain corresponding multi-layer low-level features; The multi-scale feature extraction network includes a parallel dilated convolution layer and a multi-level feature fusion module; the parallel dilated convolution layer is used to perform parallel convolution on the multi-layer low-level features by level to obtain multi-dimensional convolution features, where at least two layers are dilated convolutions with different dilation rates; the multi-level feature fusion module is used to concatenate the extracted multi-dimensional convolution features to obtain one-dimensional multi-scale features.

Images