A CSI feedback method based on multi-resolution fusion convolutional feedback network in massive MIMO system

Abstract

The present invention belongs to the technical field of intelligent mobile communication, and specifically relates to a CSI feedback method based on a multi-resolution fusion convolution feedback network in a large-scale MIMO system; the method comprises: acquiring a downlink CSI matrix in the large-scale MIMO system in real time; obtaining a CSI matrix sparse in the angle delay domain according to the CSI matrix; truncating the CSI matrix sparse in the angle delay domain to obtain a truncated matrix; processing the truncated matrix by using an MRFNet encoder to obtain a CSI compression codeword; processing the CSI compression codeword by using MRFNet to obtain a new CSI matrix; obtaining a reconstructed CSI matrix according to the new CSI matrix; the present invention realizes that a feedback network model can input CSI of different dimensional sizes, and network-fusing different environments and compression ratios, so that a mobile terminal only needs to save one set of feedback networks and can maintain a high CSI recovery accuracy.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent mobile communications, and in particular relates to a CSI feedback method based on a multi-resolution fusion convolutional feedback network in a large-scale MIMO system. Background Art

[0002] With the rapid growth of global mobile data traffic, people's demand for data transmission rate and reliability has further increased. Therefore, in order to cope with this development trend, 5G communication systems have proposed three core key technologies: ultra-dense heterogeneous cellular networks, millimeter wave communications, and large-scale multiple-input multiple-output (MIMO). Among them, large-scale MIMO refers to a communication system in which the base station uses hundreds of antennas to serve multiple user cells at the same time. Compared with the traditional MIMO system, the large-scale MIMO system can increase the capacity of the communication system several times, improve energy transmission efficiency, and reduce air interface latency. The many advantages of the above-mentioned large-scale MIMO system are based on the fact that the base station can accurately obtain the channel state information (CSI), and use precoding to eliminate interference between multiple users. However, for large-scale MIMO systems under the frequency division duplexity (FDD) standard, the uplink and downlink operate at different frequencies, so the downlink CSI is obtained by the user end and transmitted back to the base station through the feedback link. Considering that the base station uses a large number of antennas, feeding back complete CSI will result in huge resource overhead. Therefore, in order to ensure the timeliness of CSI, the feedback delay must be reduced as much as possible.

[0003] On the one hand, the traditional CSI feedback method is based on the codebook, but the number of antennas in the large-scale MIMO system is large, which increases the size and design complexity of the codebook. On the other hand, there is a CSI feedback method based on Compressed Sensing (CS), but the disadvantage of the CS feedback method is that it requires a priori assumptions about the channel, that is, it is very dependent on the sparse characteristics of the CSI matrix, but in actual scenarios, it is impossible to achieve theoretical complete sparsity; the CS recovery algorithm involves a large number of matrix iterative calculations. As the number of antennas increases, the computational complexity increases, affecting the real-time performance of the actual communication system. Therefore, the traditional CSI feedback method cannot be well applied to large-scale MIMO systems.

[0004] In recent years, as deep learning has achieved remarkable results in computer vision, natural language processing and other fields, some scholars have successfully applied it to CSI feedback. However, most of the existing CSI feedback methods based on deep learning only focus on the reconstruction accuracy of the model and ignore the practicality and computational complexity of the model. On the one hand, they rely on the fully connected layer to construct the transmission codeword, which requires the dimension of the input data to be fixed, resulting in the CSI matrix of other dimensions in practical applications cannot be applied to the existing feedback method; on the other hand, due to the limited hardware performance of the current terminal, the encoders of most feedback networks occupy a large amount of memory resources and processor resources, which cannot actually be well applied to the terminal; in summary, it is of practical significance to design a feedback network that is applicable to multiple CSI dimensions and multiple environments, has low parameter and computational complexity, and can be deployed in the terminal. Summary of the invention

[0005] In view of the shortcomings of the prior art, the present invention proposes a CSI feedback method based on a multi-resolution fusion convolutional feedback network in a large-scale MIMO system. The method comprises: acquiring a downlink CSI channel matrix in the large-scale MIMO system in real time, processing the downlink CSI channel matrix using a trained multi-resolution fusion convolutional feedback network model to obtain a new CSI channel matrix; processing the new CSI channel matrix to obtain a reconstructed CSI channel matrix;

[0006] The process of training a multi-resolution fused convolutional feedback network model includes:

[0007] S1: Get downlink CSI channel matrix;

[0008] S2: Perform a two-dimensional DFT transform on the CSI channel matrix to obtain a sparse CSI channel matrix in the angle delay domain;

[0009] S3: truncate the sparse CSI channel matrix in the angle delay domain to obtain a truncated matrix;

[0010] S4: Use a multi-resolution fusion convolutional feedback network encoder to process the truncated matrix to obtain a CSI compression codeword;

[0011] S5: Use a multi-resolution fusion convolutional feedback network decoder to process the CSI compression codeword to obtain a new CSI channel matrix;

[0012] S6: Calculate the loss function of the multi-resolution fusion convolutional feedback network model and adjust the parameters of the model according to the loss function; when the loss is minimized, the trained multi-resolution fusion convolutional feedback network model is obtained.

[0013] Preferably, the formula for performing a two-dimensional DFT transformation on the CSI channel matrix is:

[0014] H′=F c HF t H

[0015] Where H′ represents the sparse CSI channel matrix in the angle delay domain, and F c represents the left multiplication of the DFT matrix, H represents the downlink CSI channel matrix, and F t H Represents right multiplication of the DFT matrix.

[0016] Preferably, the process of processing the truncated matrix using a multi-resolution fusion convolution feedback network encoder includes: the multi-resolution fusion convolution feedback network encoder includes: two convolution layers with a convolution kernel of 3×3, three multi-resolution convolution modules MBlock and a mean pooling layer;

[0017] The first convolution layer with a convolution kernel of 3×3 is used to perform channel dimension increase on the truncated matrix to obtain the truncated matrix after channel dimension increase;

[0018] Three multi-resolution convolution modules MBlock are used to sequentially obtain the feature information of the truncated matrix after channel dimension increase at different sparsities and perform dimension reduction processing on the matrix to obtain the truncated matrix after dimension reduction.

[0019] The mean pooling layer and the second convolution layer with a convolution kernel of 3×3 are used to process the truncated matrix after dimension reduction to obtain the CSI compressed codeword.

[0020] Furthermore, the multi-resolution convolution module includes: two parallel convolution layers with convolution kernels of 3×3 and 7×7 respectively, a stacking layer and a channel shuffle layer.

[0021] Preferably, the process of using a multi-resolution fusion convolution feedback network decoder to process the CSI compression codeword includes: the multi-resolution fusion convolution feedback network decoder includes: four upsampling convolution modules UCBlock, two dense connection modules Dense Block and two convolution layers with a convolution kernel of 3×3;

[0022] Four upsampling convolution modules UCBlock are used to process the CSI compression codewords in turn to obtain the initially recovered CSI channel matrix;

[0023] The first convolution layer with a convolution kernel of 3×3 is used to perform feature extraction and channel dimension reduction processing on the initially recovered CSI channel matrix to obtain the first feature matrix;

[0024] Two densely connected modules Dense Block are used to refine the features of the first feature matrix in turn to obtain a second feature matrix;

[0025] The second convolution layer with a second convolution kernel of 3×3 is used to process the second feature matrix to obtain a feature map;

[0026] The feature map is subjected to sigmoid activation function and normalization to obtain a new CSI channel matrix.

[0027] Furthermore, the upsampling convolution module includes: an upsampling layer and a convolution layer with a convolution kernel of 3×3.

[0028] Further, using a dense connection module to refine the first feature matrix includes: the dense connection module includes three convolution layers with convolution kernels of 3×3 and one convolution layer with convolution kernels of 1×1;

[0029] Three convolutional layers with 3×3 convolution kernels are used to process the first feature matrix in sequence, where the input of each convolution layer is the concatenation of the outputs of all previous convolutional layers; the outputs of each convolution layer are concatenated to obtain a concatenated matrix;

[0030] A convolution layer with a convolution kernel of 1×1 is used to perform feature fusion and dimensionality reduction on the concatenated matrix to obtain the second feature matrix.

[0031] Preferably, calculating the loss function of the multi-resolution fusion convolutional feedback network model includes: calculating a single-task loss function of the multi-resolution fusion convolutional feedback network model; and calculating a multi-task loss function of the multi-resolution fusion convolutional feedback network model based on the single-task loss function.

[0032] Furthermore, the formula for calculating the single-task loss function is:

[0033]

[0034] Among them, L(θ) represents the single task loss, T represents the number of samples in one round of iterative training, represents the Euclidean norm, H i represents the output CSI channel matrix of the network, represents the input CSI channel matrix of the network.

[0035] Furthermore, the formula for calculating the multi-task loss function is:

[0036]

[0037] Among them, L(W) represents the multi-task loss, L η (W) represents the single-task loss when the compression ratio is η, σ η represents the noise weight in the task, and N represents the number of tasks.

[0038] The beneficial effects of the present invention are as follows: the present invention uses multi-resolution convolution to extract CSI matrix features with different sparsities in view of the fact that CSI in different environments has different sparsities, so that the feedback network model can input CSI of different dimensional sizes, thereby increasing the generalization ability of the feedback network; depthwise separable convolution is selected to replace ordinary convolution to solve the problem of increased complexity caused by multi-resolution convolution; the convolution layer is used to realize the function of the fully connected layer, so that the feedback network avoids practical problems caused by the use of the fully connected layer; different environments and compression ratios are integrated into the network, so that the mobile terminal only needs to save one set of feedback networks and can maintain a high CSI recovery accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is a flow chart of the CSI feedback method based on the multi-resolution fusion convolution feedback network in the present invention;

[0040] Figure 2 It is a schematic diagram of the MRFNet feedback network structure in the present invention;

[0041] Figure 3 It is a schematic diagram of the DenseBlock structure in the MRFNet decoder of the present invention;

[0042] Figure 4 This is a comparison chart of the NMSE of the present invention with the CsiNet feedback network at different compression ratios in an indoor environment;

[0043] Figure 5 This is a comparison chart of the NMSE of the present invention and the CsiNet feedback network under different compression ratios in an outdoor environment. DETAILED DESCRIPTION

[0044] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0045] The present invention proposes a CSI feedback method based on a multi-resolution fusion convolutional feedback network in a large-scale MIMO system. Figure 1 As shown, the method includes: acquiring a downlink CSI channel matrix in a massive MIMO system in real time, processing the downlink CSI channel matrix using a trained multi-resolution fusion convolution feedback network model to obtain a new CSI channel matrix; processing the new CSI channel matrix to obtain a reconstructed CSI channel matrix;

[0046] The process of training a multi-resolution fused convolutional feedback network model includes:

[0047] S1: Get downlink CSI channel matrix;

[0048] S2: Perform a two-dimensional DFT transform on the CSI channel matrix to obtain a sparse CSI channel matrix in the angle delay domain;

[0049] S3: truncate the sparse CSI channel matrix in the angle delay domain to obtain a truncated matrix;

[0050] S4: Use a multi-resolution fusion convolutional feedback network encoder to process the truncated matrix to obtain a CSI compression codeword;

[0051] S5: Use a multi-resolution fusion convolutional feedback network decoder to process the CSI compression codeword to obtain a new CSI channel matrix;

[0052] S6: Calculate the loss function of the multi-resolution fusion convolutional feedback network model and adjust the parameters of the model according to the loss function; when the loss is minimized, the trained multi-resolution fusion convolutional feedback network model is obtained.

[0053] Considering the massive MIMO system under FDD standard, the base station antenna adopts a uniform linear array (ULA) with half-wavelength spacing, and the number of transmitting antennas at the base station is N. t >>1, considering the single user case, the number of receiving antennas is N r =1. Assume that OFDM has N c orthogonal subcarriers, where the signal vector received by the nth subcarrier can be expressed as:

[0054] y n =Ax n +z n

[0055] in, represents the received vector of the nth OFDM symbol period, represents the transmit vector of the nth OFDM symbol period, represents additive white Gaussian noise, A represents the diagonal channel matrix represents the channel frequency domain vector on the i-th subcarrier, Represents the precoding vector on the i-th subcarrier.

[0056] A preferred embodiment of training a multi-resolution fusion convolutional feedback network model is as follows:

[0057] In the downlink of the massive MIMO system, the COST 2100 model was used to set up two experimental environments: an indoor scenario with a signal frequency of 5.3 GHz and an outdoor environment with a signal frequency of 300 MHz. In the indoor scenario, the base station was located at the center of a square area with a side length of 20 meters, and in the outdoor scenario, the base station was located at the center of a square area with a side length of 400 meters. The user terminals in the indoor or outdoor scenarios were randomly placed in their respective square areas. The base station used N t = 32 transmit antennas, and each user end uses a single receive antenna. This massive MIMO system uses OFDM carrier modulation, using N c =1024 subcarriers. Based on the above conditions, 120,000 data sets are generated, of which 100,000 data sets are randomly selected for training. The cross-validation method is used to divide the training set and validation set of the model in a ratio of 8:2. The remaining 20,000 data sets are used as test sets to evaluate the CSI reconstruction accuracy of the model.

[0058] The base station obtains the channel response matrix h i , according to the channel response matrix h i The downlink CSI channel matrix H is obtained as:

[0059]

[0060] Expand the downlink CSI channel matrix H:

[0061]

[0062] Each column vector of the above matrix H is a frequency domain channel vector. DFT transformation is performed on each column vector, that is, the dimension of H multiplied on the left is N c ×N c The DFT matrix of the delay domain is obtained. Due to the finite delay spread, only the first N a (N a <N c ) rows have non-zero values; each row vector in H is a time domain channel vector, and a DFT transform is performed on each row vector of H, that is, the dimension of H right multiplication is N t ×N t The DFT matrix of the spatial frequency domain is transformed into the angle domain channel vector; that is, the sparse CSI matrix in the angle delay domain can be obtained by performing a two-dimensional DFT transformation on the CSI matrix H in the spatial frequency domain. The process can be expressed as:

[0063] H′=F c HF t H

[0064] Wherein, H′ represents the sparse CSI channel matrix in the angle delay domain obtained after two-dimensional DFT transformation, H represents the downlink CSI channel matrix, preferably, F c F represents the left multiplication of the DFT matrix, the size is 1024×1024; t H Represents the right multiplication of the DFT matrix, the size of which is 32×32.

[0065] Due to the multipath effect, the delay between arrival times is within a limited time range, so in the delay domain, the CSI matrix is ​​only a (N a <N c ) rows have non-zero values, and the values ​​of the remaining rows are all 0 or close to 0, that is, the CSI matrix H′ in the angle delay domain is sparse; the first N rows of the CSI matrix H′ are a The rows are truncated to obtain the truncated matrix H a , preferably, set N a =32, that is, the first 32 rows of H′ are truncated to obtain the truncated matrix H a .

[0066] Since the current deep learning framework does not support complex number calculations, it is necessary to convert the input CSI matrix, i.e., the truncated matrix H a The real and imaginary parts of the complex number are split and then concatenated in the channel dimension. After concatenation, the CSI matrix dimension is 2×32×32.

[0067] Construct a multi-resolution fusion convolution feedback network (Multi-Resolution and Fusion Network, MRFNet) encoder, which belongs to the user side. Its network structure is as follows Figure 2 As shown in the Encoder section, the encoder consists of two convolutional layers with a convolution kernel of 3×3, three multi-resolution convolutional modules MBlock and a mean pooling layer.

[0068] Use MRFNet encoder to truncate the matrix H a The specific process is as follows:

[0069] The first convolution layer with a convolution kernel of 3×3 is used to perform channel dimension increase on the truncated matrix to obtain the truncated matrix after channel dimension increase.

[0070] Three multi-resolution convolution modules MBlock are used to successively obtain the feature information of the truncated matrix after channel dimension increase at different sparsities and perform dimension reduction processing on the matrix; wherein, the multi-resolution convolution module includes two parallel convolution layers with convolution kernels of 3×3 and 7×7 respectively, a superposition layer and a channel shuffle layer. The parallel convolution layer is used to extract the feature information in the CSI at different sparsities and at the same time increase the channel dimension of the input CSI matrix and reduce the dimension of the CSI matrix. In order to avoid the problem of large computational complexity caused by multi-resolution convolution, multi-resolution convolution uses depthwise separable convolution. The superposition layer is used to merge the channel dimensions of different matrices. The channel shuffle layer rearranges and combines the feature matrices on the channel dimension, so that the feature information between different channels is integrated with each other, thereby increasing the feature extraction capability of the multi-resolution convolution module; the truncated matrix after channel dimension increase is processed by three MBlocks to obtain the truncated matrix after dimension reduction.

[0071] The truncated matrix after dimension reduction is processed by the mean pooling layer and the second convolution layer with a convolution kernel of 3×3 to obtain the CSI compressed codeword. The dimension of the compressed codeword is determined by the compression ratio. If the compression ratio is 1 / 4, the output codeword dimension is Among them, zero padding, LeakyReLU activation function and batch normalization are used in the above convolution calculations.

[0072] Construct the MRFNet decoder, whose network structure is as follows Figure 2 As shown in the Decoder section, the decoder consists of four upsampling convolution modules UCBlock, two densely connected modules Dense Block, and two convolution layers with a convolution kernel of 3×3.

[0073] The CSI compressed codeword output by the encoder is input into the decoder of MRFNet for processing. The specific processing process is:

[0074] Four upsampling convolution modules UCBlock are used to process the CSI compressed codewords in sequence. The upsampling convolution module includes an upsampling layer and a convolution layer with a convolution kernel of 3 × 3. After the CSI compressed codewords are processed by four upsampling layers, the CSI compressed codewords are restored to the same size as the original CSI, and the initially restored CSI channel matrix is ​​obtained, with a size of 64 × 32 × 32.

[0075] The first convolution layer with a convolution kernel of 3×3 is used to perform feature extraction and channel dimension reduction on the initially recovered CSI channel matrix to obtain the first feature matrix with a size of 2×32×32.

[0076] Two densely connected modules Dense Block are used to refine the first feature matrix in turn, and each Dense Block outputs a 2×32×32 feature matrix; Figure 3 As shown in the figure, the dense connection module includes three convolution layers with convolution kernels of 3×3 and one convolution layer with convolution kernels of 1×1; the process of processing the feature matrix by the dense connection module Dense Block is as follows: three convolution layers with convolution kernels of 3×3 are used to process the feature matrix in turn, where the input of each convolution layer is the concatenation of the outputs of all previous convolution layers; the output of each convolution layer is concatenated to obtain a concatenated matrix; the concatenated matrix is ​​input into a convolution layer with a convolution kernel of 1×1 to perform feature fusion and dimensionality reduction to obtain a second feature matrix.

[0077] The second feature matrix output by the second densely connected module is processed by a second convolution layer with a second convolution kernel of 3×3 to obtain a feature map with a size of 2×32×32; wherein, zero padding, LeakyReLU activation function and batch normalization are used in the above convolution calculations.

[0078] The feature map is subjected to sigmoid activation function and normalization processing, and a real number matrix of 2×32×32 is output, which is used as the real and imaginary parts of the new CSI channel matrix.

[0079] The single-task loss function of MRFNet is defined as the mean square error, and its formula is as follows:

[0080]

[0081] Among them, L(θ) represents the single task loss, T represents the number of samples in one round of iterative training, represents the Euclidean norm, H i represents the output CSI channel matrix of the network, represents the input CSI channel matrix of the network.

[0082] The multi-task loss function of the multi-resolution fusion convolutional feedback network can be calculated based on the single-task loss function. The derivation process is as follows:

[0083] Let f(x) be the output of the neural network and define the likelihood as a Gaussian distribution. The probability model is expressed as:

[0084]

[0085] Here, σ represents the standard deviation of the noise.

[0086] The probability model under multi-task can be expressed as:

[0087]

[0088] where y1,…y n Represents the output of the neural network corresponding to different tasks.

[0089] Logarithmization of the above formula can give the log-likelihood, which is:

[0090]

[0091] Assuming that the neural network has two different task outputs y1 and y2, according to the above formula, the multi-task loss function is:

[0092]

[0093] make Substituting the above formula into the multi-task loss function is:

[0094]

[0095] Among them, L(W) represents the multi-task loss, L η (W) represents the single-task loss when the compression ratio is η, σ η Represents the noise weight in the task. The greater the noise, the smaller the loss weight of the task. N represents the number of tasks.

[0096] During the training process, the feedback network model parameters are initialized using the Xavier method, the model training optimizer uses the Adam algorithm, and the model learning rate is dynamically adjusted using the cosine annealing algorithm. Each iteration uses 200 samples in the training set to calculate the gradient, and updates the parameters according to the formula of the Adam algorithm. In this way, the entire training set is trained 1,000 times, and the model parameters are debugged according to the results of the validation set. When the loss function is minimized, the training is stopped, and the performance of the final model is verified according to the test set results.

[0097] The downlink CSI channel matrix of the massive MIMO system is obtained in real time, and the downlink CSI channel matrix is ​​compressed and reconstructed using the trained multi-resolution fusion convolutional feedback network model. A zero padding operation is performed to restore the original CSI matrix H to the same size, and then a two-dimensional inverse DFT transform is performed to obtain the reconstructed value of the original space-frequency domain CSI matrix, that is, the reconstructed CSI channel matrix.

[0098] The present invention was evaluated by using the same data set to train the existing deep learning-based CsiNet feedback network and comparing it with the present invention. The results are as follows:

[0099] Performance comparison of several feedback networks under different compression ratios in indoor and outdoor environments Figure 4 and Figure 5As shown in the figure, MRNet is a multi-resolution single-task feedback network, MRFNet-CR represents a multi-task feedback network that integrates four compression ratios under indoor or outdoor dataset training, and MRFNet-CRE represents a multi-task feedback network that integrates four compression ratios under indoor and outdoor mixed dataset training. The performance between different feedback networks is expressed in the logarithmic form of NMSE. The model complexity comparison between MRFNet and CsiNet under a single task is shown in Table 1, expressed in floating-point parameter quantity and floating-point calculation quantity.

[0100] Table 1 Comparison of encoder complexity of different models

[0101]

[0102] As can be seen from Table 1, the number of parameters of the MRFNet encoder is smaller than that of CsiNet when the compression ratio is 1 / 4 and 1 / 8, but larger than that of CsiNet when the compression ratio is 1 / 16 and 1 / 32. The reason is that 99% of the parameters of the CsiNet encoder are distributed in the fully connected layer, and the number of parameters of the fully connected layer will decrease rapidly as the compression ratio increases. The number of parameters of MRFNet is mainly distributed in the front multi-resolution convolutional layer. Therefore, as the compression ratio increases, the number of parameters does not decrease as much as CsiNet. However, when CsiNet copes with four different compression ratios in two environments, the total number of parameters of its encoder is about 3934k, which requires about 15MB of hardware memory, while MRFNet only needs one set of feedback network, and the total number of parameters of its encoder is about 1305k, which requires about 5MB of hardware memory. Compared with CsiNet, the consumption of hardware memory is reduced by about 66%. In terms of computational complexity, MRFNet is much higher than CsiNet. When the compression ratio is 1 / 4, the computational complexity is about 10 times that of CsiNet, and when the compression ratio is 1 / 32, the computational complexity is about 56 times that of CsiNet. The reason is that the computational complexity of the fully connected layer in CsiNet accounts for 90% of the total computational complexity of the entire encoder. Therefore, as the compression ratio increases, the computational complexity will decrease rapidly with the increase of the compression ratio, while the computational complexity of MRFNet is mainly concentrated in the multi-resolution convolutional layer in the first half, so it will not be significantly reduced as the compression ratio increases. Although the computational complexity of MRFNet is much higher than that of CsiNet, it can still be applied to existing mobile terminals. For example, Kirin 970 is currently a mid-range mobile SoC chip with a maximum computing power of 244.8G, while the maximum computing complexity of MRFNet is 11493k, which requires about 11G of computing power, about 4% of the total computing power of the processor, which is completely acceptable for the terminal.

[0103] Depend on Figure 4It can be seen that in indoor environments, the NMSE of the feedback network proposed in this paper is higher than that of CsiNet under four different compression ratios, regardless of whether multi-task fusion is performed. When the compression ratio is 1 / 4 and 1 / 8, the NMSE performance of MRNet is almost the same as that of MRFNet-CR, while MRFNet-CR is slightly better than MRNet when the compression ratio is 1 / 16 and 1 / 32. As the compression ratio increases, the difference between the NMSE of MRFNet-CR and MRNet gradually increases. This is because multi-task learning can effectively average the noise in the data, thereby avoiding the overfitting phenomenon caused by the single-task model learning the noise. When the feedback network chooses a large compression ratio, the data contains less effective information. At this time, the feedback network can use some features learned at a small compression ratio to assist in feature learning. The single-task model may be more likely to learn some noise, resulting in a decrease in performance. In addition, the indoor and outdoor environments are fused based on the fusion of the multi-compression ratio network. It can be seen from the figure that the feedback performance of MRFNet-CRE has dropped significantly. The reason is that, unlike the multi-task training process that relies on the weights of the loss function for adjustment, the feedback network is directly trained through the data sets of the two environments. The feedback network will choose to balance between the two environments. Therefore, compared with the feedback network trained in a single environment, its feedback performance has dropped significantly. However, thanks to the multiple multi-resolution convolution modules in the encoder, the feedback network is provided with more opportunities to learn different features at the same time, so it still has higher feedback accuracy than CsiNet.

[0104] Depend on Figure 5 It can be seen that in outdoor environments, the NMSE of the feedback network proposed in this paper is higher than that of CsiNet under four different compression ratios, regardless of whether multi-task fusion is performed. Under four different compression ratio conditions, the NMSE of MRFNet-CR is better than that of MRNet, and as the compression ratio increases, the difference between the NMSE of MRFNet-CR and MRNet gradually increases. This is because the CSI in the outdoor environment is more complex than that in the indoor environment, and the data contains more noise. Therefore, after the fusion of multiple compression ratio networks, the influence of noise on model training under each compression ratio is weakened, and the common features learned by the common part of the multi-task model have a more obvious gain for each task than the indoor environment. Similarly, the indoor and outdoor environments are fused on the basis of the fusion of multiple compression ratio networks. It can be seen from the figure that, like the indoor environment, the feedback performance of MRFNet-CRE has dropped significantly, and the reason is roughly the same as that of the indoor environment. Thanks to the multiple multi-resolution convolution modules in the encoder, it still has higher feedback accuracy than CsiNet in the outdoor environment.

[0105] In summary, although the computational complexity of the present invention is higher than that of the CsiNet feedback network, it is within the acceptable range of the terminal, and the CSI feedback accuracy of the present invention in both indoor and outdoor environments is higher than that of the CsiNet feedback network.

[0106] Different from the existing CSI feedback network model based on deep learning, which requires the input CSI dimension size to be fixed, the present invention uses multi-resolution convolution to extract CSI matrix features with different sparsities based on the characteristics that CSI in different environments has different sparsities, so that the feedback network model can input CSI of different dimensional sizes, thereby increasing the generalization ability of the feedback network; deep separable convolution is selected to replace ordinary convolution to solve the problem of increased complexity brought by multi-resolution convolution; the present invention uses convolution layer to realize the function of fully connected layer, so that the feedback network avoids practical problems caused by the use of fully connected layer; different from the existing CSI feedback network model based on deep learning, in which each environment needs to correspond to a set of network models respectively, the present invention integrates different environments and compression ratios into a network, so that the mobile terminal only needs to save one set of feedback network, which can be applied to a variety of environments and can maintain a high CSI recovery accuracy.

[0107] The above embodiments further illustrate the purpose, technical solutions and advantages of the present invention in detail. It should be understood that the above embodiments are only preferred implementation modes of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made to the present invention within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A CSI feedback method based on a multi-resolution fusion convolutional feedback network in a large-scale MIMO system, characterized in that: include: Acquire the downlink CSI channel matrix in the massive MIMO system in real time, use the trained multi-resolution fusion convolution feedback network model to process the downlink CSI channel matrix to obtain a new CSI channel matrix; process the new CSI channel matrix to obtain a reconstructed CSI channel matrix; The process of training a multi-resolution fused convolutional feedback network model includes: S1: Get downlink CSI channel matrix; S2: Perform a two-dimensional DFT transform on the CSI channel matrix to obtain a sparse CSI channel matrix in the angle delay domain; S3: truncate the sparse CSI channel matrix in the angle delay domain to obtain a truncated matrix; S4: The truncated matrix is ​​processed by a multi-resolution fusion convolution feedback network encoder to obtain a CSI compression codeword; the process of processing the truncated matrix by a multi-resolution fusion convolution feedback network encoder includes: The multi-resolution fusion convolution feedback network encoder includes: two convolution layers with 3×3 convolution kernels, three multi-resolution convolution modules MBlock and a mean pooling layer; the multi-resolution convolution module includes: two parallel convolution layers with 3×3 and 7×7 convolution kernels, a stacking layer and a channel shuffle layer; The first convolution layer with a convolution kernel of 3×3 is used to perform channel dimension increase on the truncated matrix to obtain the truncated matrix after channel dimension increase; Three multi-resolution convolution modules MBlock are used to sequentially obtain the feature information of the truncated matrix after channel dimension increase at different sparsities and perform dimension reduction processing on the matrix to obtain the truncated matrix after dimension reduction. The truncated matrix after dimension reduction is processed by the mean pooling layer and the second convolution layer with a convolution kernel of 3×3 to obtain the CSI compressed codeword; S5: Processing the CSI compression codeword using a multi-resolution fusion convolution feedback network decoder to obtain a new CSI channel matrix; the process of processing the CSI compression codeword using a multi-resolution fusion convolution feedback network decoder includes: The multi-resolution fusion convolution feedback network decoder includes: four upsampling convolution modules UCBlock, two densely connected modules Dense Block and two convolution layers with a convolution kernel of 3×3; the four upsampling convolution modules UCBlock are used to process the CSI compression codewords in turn to obtain a preliminarily restored CSI channel matrix; the upsampling convolution module includes: an upsampling layer and a convolution layer with a convolution kernel of 3×3; the first convolution layer with a convolution kernel of 3×3 is used to perform feature extraction and channel dimension reduction on the preliminarily restored CSI channel matrix to obtain the first feature matrix; Two densely connected modules Dense Block are used to refine the features of the first feature matrix in turn to obtain a second feature matrix; the densely connected module is used to refine the first feature matrix, including: the densely connected module includes three convolution layers with convolution kernels of 3×3 and one convolution layer with convolution kernels of 1×1; three convolution layers with convolution kernels of 3×3 are used to process the first feature matrix in turn, wherein the input of each convolution layer is the concatenation of the outputs of all previous convolution layers; the outputs of each convolution layer are concatenated to obtain a concatenated matrix; a convolution layer with a convolution kernel of 1×1 is used to perform feature fusion and dimensionality reduction on the concatenated matrix to obtain a second feature matrix; The second convolution layer with a second convolution kernel of 3×3 is used to process the second feature matrix to obtain a feature map; Perform sigmoid activation function and normalization processing on the feature map to obtain a new CSI channel matrix; S6: Calculate the loss function of the multi-resolution fusion convolutional feedback network model and adjust the parameters of the model according to the loss function; when the loss is minimized, the trained multi-resolution fusion convolutional feedback network model is obtained.

2. According to claim 1, a CSI feedback method based on a multi-resolution fusion convolutional feedback network in a large-scale MIMO system is characterized in that: The formula for performing two-dimensional DFT transformation on the CSI channel matrix is: H′=F c HF t H Where H′ represents the sparse CSI channel matrix in the angle delay domain, and F c represents the left multiplication of the DFT matrix, H represents the downlink CSI channel matrix, and F t H Represents right multiplication of the DFT matrix.

3. The CSI feedback method based on a multi-resolution fusion convolutional feedback network in a large-scale MIMO system according to claim 1, characterized in that: Calculating the loss function of the multi-resolution fusion convolutional feedback network model includes: calculating the single-task loss function of the multi-resolution fusion convolutional feedback network model; and calculating the multi-task loss function of the multi-resolution fusion convolutional feedback network model according to the single-task loss function.

4. The CSI feedback method based on a multi-resolution fusion convolutional feedback network in a large-scale MIMO system according to claim 3, characterized in that: The formula for calculating the single-task loss function is: Among them, L(θ) represents the single task loss, T represents the number of samples in one round of iterative training, represents the Euclidean norm, H i represents the output CSI channel matrix of the network, represents the input CSI channel matrix of the network.

5. The CSI feedback method based on a multi-resolution fusion convolutional feedback network in a large-scale MIMO system according to claim 3, characterized in that: The formula for calculating the multi-task loss function is: Among them, L(W) represents the multi-task loss, L η (W) represents the single-task loss when the compression ratio is η, σ η represents the noise weight in the task, and N represents the number of tasks.

Images