A method and apparatus for encoding and decoding channel state information (CSI)

Abstract

The present application proposes a CSI encoding and decoding method and device, which relates to the field of wireless communication technology. The encoding scheme is: the UE encodes the target CSI matrix in the delay angle domain based on the first target CSI compression encoder to generate a compressed coding value. The first target CSI compression encoder includes N compound convolution layers and a fully connected layer. Each compound convolution layer includes a delay domain convolution step and an angle domain convolution step. The delay domain convolution step of the first compound convolution layer is smaller than the angle domain convolution step of the first compound convolution layer, and N is a positive integer. In the present application, since the correlation of the target CSI matrix is strong in the angle domain and low in the delay domain, the first compound convolution layer can be used to compress the angle domain dimension with strong correlation more and compress the delay domain dimension with low correlation less, which not only improves the compression performance of the encoder, but also realizes reasonable compression coding of the target CSI matrix.

Description

Technical Field

[0001] The present application relates to the field of mobile communications, and in particular to a CSI encoding and decoding method and device. Background Art

[0002] Massive Multiple-Input Multiple-Output (m-MIMO) antenna systems have become an essential component of wireless communication networks. This m-MIMO technology requires transmitters to obtain accurate channel state information (CSI). In practice, the user equipment (UE) estimates the downlink CSI and then feeds it back to the network. However, due to the large number of antennas, the channel matrix in an m-MIMO antenna system is very large, making CSI estimation and feedback very challenging, especially over bandwidth-limited feedback channels. Summary of the Invention

[0003] The channel state information (CSI) encoding and decoding method and apparatus, communication equipment, and storage medium proposed in this application are used to compress and encode CSI information in a UE and decode it in a network device.

[0004] The first embodiment of the present application provides a CSI encoding method applicable to a UE, the method including:

[0005] The target CSI matrix in the delay angle domain is encoded based on the first target CSI compression encoder to generate a compressed coding value, wherein the first target CSI compression encoder includes N compound convolutional layers and one fully connected layer, each compound convolutional layer includes a delay domain convolution step and an angle domain convolution step, the delay domain convolution step of the first compound convolution layer among the N compound convolution layers is smaller than the angle domain convolution step of the first compound convolution layer, and N is a positive integer.

[0006] The second aspect embodiment of the present application proposes a CSI decoding method, which is applicable to a network device, and the method includes: receiving a compressed coding value corresponding to a target CSI matrix in the delay angle domain sent by a UE; decoding the compressed coding value based on a first target CSI decoder to generate an estimated value of the target CSI matrix, wherein the first target CSI decoder includes N compound deconvolution layers and a fully connected layer, each compound deconvolution layer includes a delay domain convolution step and an angle domain convolution step, the delay domain convolution step of the Nth compound deconvolution layer among the N compound deconvolution layers is smaller than the angle domain convolution step of the Nth compound deconvolution layer, and N is a positive integer.

[0007] An embodiment of the third aspect of the present application proposes a CSI encoding device, which includes: an encoding module for encoding a target CSI matrix in the delay angle domain based on a first target CSI compression encoder to generate a compressed coding value, wherein the first target CSI compression encoder includes N compound convolutional layers and a fully connected layer, each compound convolutional layer includes a delay domain convolution step and an angle domain convolution step, the delay domain convolution step of the first compound convolution layer among the N compound convolution layers is smaller than the angle domain convolution step of the first compound convolution layer, and N is a positive integer.

[0008] An embodiment of the fourth aspect of the present application proposes a CSI decoding device, which includes: a receiving module for receiving a compressed coding value corresponding to a target CSI matrix in the delay angle domain sent by a UE; a decoding module for decoding the compressed coding value based on a first target CSI decoder to generate an estimated value of the target CSI matrix, wherein the first target CSI decoder includes N composite deconvolution layers and a fully connected layer, each composite deconvolution layer includes a delay domain convolution step and an angle domain convolution step, the delay domain convolution step of the Nth composite deconvolution layer among the N composite deconvolution layers is smaller than the angle domain convolution step of the Nth composite deconvolution layer, and N is a positive integer.

[0009] The fifth aspect embodiment of the present application provides a communication device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the CSI encoding method described in the first aspect embodiment of the present application, or the CSI decoding method described in the second aspect embodiment of the present application.

[0010] The sixth aspect embodiment of the present application provides a computer storage medium, wherein the computer storage medium stores computer-executable instructions. After the computer-executable instructions are executed by a processor, the CSI encoding method described in the first aspect embodiment of the present application or the CSI decoding method described in the second aspect embodiment of the present application can be implemented.

[0011] It should be understood that the content described in this application is not intended to identify the key or important features of the embodiments of this application, nor is it intended to limit the scope of this application. Other features of this application will become easy to understand through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] The above and / or additional aspects and advantages of the present application will become apparent and easily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0013] Figure 1A schematic diagram of a CSI encoding method provided in an embodiment of the present application;

[0014] Figure 2 A schematic diagram of the structure of the first target CSI compression encoder provided in an embodiment of the present application;

[0015] Figure 3 A schematic diagram of a flow chart of another CSI encoding method provided in an embodiment of the present application;

[0016] Figure 4 A schematic diagram of a flow chart of another CSI encoding method provided in an embodiment of the present application;

[0017] Figure 5 A schematic diagram of a CSI decoding method provided in an embodiment of the present application;

[0018] Figure 6 This is a schematic structural diagram of a first target CSI decoder provided in an embodiment of the present application;

[0019] Figure 7 A flowchart of another CSI decoding method provided in an embodiment of the present application;

[0020] Figure 8 A schematic diagram of a CSI transmission method provided in an embodiment of the present application;

[0021] Figure 9 A schematic diagram of a flow chart of another CSI transmission method provided in an embodiment of the present application;

[0022] Figure 10 A schematic diagram of a flow chart of another CSI transmission method provided in an embodiment of the present application;

[0023] Figure 11 A schematic diagram of the structure of a CSI encoding device provided in an embodiment of the present application;

[0024] Figure 12 A schematic diagram of the structure of another CSI encoding device provided in an embodiment of the present application;

[0025] Figure 13 A schematic diagram of the structure of a CSI decoding device proposed in an embodiment of the present application;

[0026] Figure 14 A schematic diagram of the structure of another CSI decoding device proposed in an embodiment of the present application;

[0027] Figure 15 A schematic diagram of the structure of a communication device proposed in an embodiment of the present application. DETAILED DESCRIPTION

[0028] The following describes in detail embodiments of the present application, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements, or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present application, and are not to be construed as limiting the present application.

[0029] Figure 1 The present invention provides a flow chart of a CSI encoding method. The CSI encoding method is executed by the UE. Figure 1 As shown, the CSI encoding method includes the following steps:

[0030] S101, encode the target CSI matrix in the delay angle domain based on the first target CSI compression encoder to generate a compressed coding value, wherein the first target CSI compression encoder includes N compound convolutional layers and one fully connected layer, each compound convolutional layer includes a delay domain convolution step and an angle domain convolution step, the delay domain convolution step of the first compound convolution layer among the N compound convolution layers is smaller than the angle domain convolution step of the first compound convolution layer, and N is a positive integer.

[0031] Figure 2 This is a schematic diagram of the structure of the first target CSI compression encoder provided in the embodiment of the present application. Figure 2 As shown, the first target CSI compression encoder includes N composite convolutional layers and one fully connected layer. In the implementation, each of the N composite convolutional layers includes a convolutional layer, a batch normalization layer, and an activation function layer.

[0032] It should be noted that the convolutional layers within each composite convolutional layer use a convolution kernel of size f × h × w × c, with a convolution step of (g, p), where f represents the number of convolution kernels, h represents the height of the convolution kernel, w represents the width of the convolution kernel, c represents the number of output channels of the convolution kernel, g represents the convolution step in the time domain, and p represents the convolution step in the angle domain. f is an adjustable parameter that can be adjusted according to usage requirements. f, h, w, c, g, and p are all positive integers.

[0033] Optionally, the first composite convolutional layer uses a convolutional layer with a kernel size of f×1×m×n and a convolution stride of (k, l), where f corresponds to the number of kernels, 1 corresponds to the kernel height, m corresponds to the kernel width, n corresponds to the number of kernel input channels, k corresponds to the convolution stride in the time domain, and l corresponds to the convolution stride in the angle domain. f is an adjustable parameter that can be adjusted based on usage requirements. f, m, n, k, and l are all positive integers.

[0034] Because the target CSI matrix in the delay angle domain has different correlations in the delay and angle domains—that is, the correlation is stronger in the angle domain and lower in the delay domain—in order to properly compress and encode the target CSI matrix, in this application example, the first composite convolution layer in the first target CSI compression encoder is required to ensure that the delay domain convolution step size k is smaller than the angle domain convolution step size l. This restriction allows for greater compression of the highly correlated angle domain dimension and less compression of the less correlated delay domain dimension, resulting in better compression performance for the encoder.

[0035] In this embodiment of the present application, the first composite convolutional layer also needs to meet the following conditions:

[0036] The delay domain convolution step size of the first composite convolution layer is smaller than the width of the convolution kernel in the first composite convolution layer. That is, the delay domain convolution step size k in the first composite convolution layer is smaller than the width m of the convolution kernel.

[0037] The angular domain convolution step size of the first composite convolution layer is smaller than the number of input channels of the convolution kernel in the first composite convolution layer. In other words, the angular domain convolution step size l in the first composite convolution layer is smaller than the width n of the convolution kernel.

[0038] UE converts the CSI matrix H in the space-frequency domain f Perform a two-dimensional discrete Fourier transform (DFT) to obtain a target CSI matrix in the delay angle domain. Further, the target CSI matrix is input into a first target CSI compression encoder, and the first target CSI compression encoder compresses and encodes the target CSI matrix to obtain a compressed coding value. Optionally, the real part and / or imaginary part of the target CSI matrix is input into the first composite convolution layer in the first target CSI compression encoder to obtain a tensor output by the first composite convolution layer. In some implementations, the real part or imaginary part of the target CSI matrix is input into the first composite convolution layer, respectively, or the real part and imaginary part of the target CSI matrix are input into the first composite convolution layer in sequence.

[0039] Furthermore, the tensor output by the previous composite convolution layer is used as the input of the next composite convolution layer until the Nth composite convolution layer. Finally, the tensor output by the Nth composite convolution layer is reconstructed, and the reconstructed tensor is input into the fully connected layer, which performs a fully connected operation to generate a compressed coding value of the target CSI matrix. It should be noted that the size of the tensor output by the previous composite convolution layer may change under the influence of the convolution step size of the next composite convolution layer after the convolution process. Taking the first composite convolution layer as an example, the size of the target CSI matrix is 1×N cc ×N t, input it to the first composite convolutional layer, and the size of the output tensor after convolution is f×N cc / k×N t / l.

[0040] Optionally, in a scenario where the UE deploys only the first target CSI compression encoder, since the real and imaginary parts of the CSI matrix have similar correlations, the first target CSI compression encoder can be used to compress and encode the real and imaginary parts of the target CSI matrix. In some implementations, the UE can extract the real and imaginary parts of the target CSI matrix and input the real and imaginary parts of the target CSI matrix into the first target CSI compression encoder in sequence for encoding. In other words, the UE serially inputs the real and imaginary parts of the target CSI matrix into the first target CSI compression encoder for encoding. In the serial encoding method, since only one target CSI compression encoder is deployed, the serial encoding method occupies fewer resources.

[0041] Since the correlation between the real part and the imaginary part of the CSI matrix is similar, the parameters of the compression encoders used when encoding the real part and the imaginary part are often similar or the same. Therefore, optionally, based on the model parameters of the first target CSI compression encoder, a second target CSI compression encoder is constructed to extract the real part and the imaginary part of the target CSI matrix, and the real part and the imaginary part of the target CSI matrix are respectively input into the corresponding target CSI compression encoders for encoding. In other words, the UE deploys two target CSI compression encoders at the same time, and can input the real part of the target CSI matrix into one of the target CSI compression encoders, and input the imaginary part of the target CSI matrix into the other target CSI compression encoder in parallel. In the embodiment of the present application, in the parallel coding method, since two target CSI compression encoders are deployed at the same time, the real part and the imaginary part of the target CSI matrix can be encoded synchronously, so that the encoding speed of the parallel coding method is faster.

[0042] It can be understood that if the real part corresponds to the first target CSI compression encoder, the imaginary part corresponds to the second target CSI compression encoder, or, if the real part corresponds to the second target CSI compression encoder, the imaginary part corresponds to the first target CSI compression encoder. Optionally, if the real part of the sample CSI matrix is used to train the first target CSI compression encoder, then the real part of the target CSI matrix corresponds to the first target CSI compression encoder, and the imaginary part of the target CSI matrix corresponds to the first target CSI compression encoder. If the imaginary part of the sample CSI matrix is used to train the first target CSI compression encoder, then the imaginary part of the target CSI matrix corresponds to the first target CSI compression encoder, and the real part of the target CSI matrix corresponds to the second target CSI compression encoder.

[0043] In the CSI encoding method provided in the embodiment of the present application, the delay domain convolution step size of the first composite convolution layer in the first target CSI compression encoder deployed on the UE is smaller than the angle domain convolution step size. Since the correlation of the target CSI matrix in the angle domain is strong and the correlation in the delay domain is low, the first composite convolution layer can perform more compression on the angle domain dimension with strong correlation and less compression on the delay domain dimension with low correlation, which not only improves the compression performance of the encoder but also realizes reasonable compression encoding of the target CSI matrix.

[0044] Figure 3 The present invention provides another CSI encoding method according to an embodiment of the present invention. The CSI encoding method is executed by the UE. Figure 3 As shown, the CSI encoding method includes the following steps:

[0045] S301 : Training a CSI compression encoder based on the real part or the imaginary part of the sample CSI matrix in the delay angle domain to obtain a first target CSI compression encoder.

[0046] In order to implement CSI compression coding, it is necessary to pre-train a first target CSI compression encoder, and at the beginning of training, randomly initialize the parameters of each layer. In an embodiment of the present application, an initial CSI compression encoder is pre-constructed, and the CSI compression encoder includes N compound convolutional layers and a fully connected layer, wherein each compound convolutional layer includes a delay domain convolution step and an angle domain convolution step, and the delay domain convolution step of the first compound convolutional layer among the N compound convolutional layers is smaller than the angle domain convolution step of the first compound convolutional layer, and N is a positive integer.

[0047] Optionally, the first composite convolutional layer also needs to meet the following conditions:

[0048] The delay domain convolution step size of the first composite convolution layer is smaller than the width of the convolution kernel in the first composite convolution layer;

[0049] The angle domain convolution stride of the first composite convolution layer is smaller than the number of input channels of the convolution kernel in the first composite convolution layer.

[0050] In an embodiment of the present application, the UE may collect a sample CSI matrix in the delay angle domain. Optionally, the UE collects a sample CSI matrix in the space-frequency domain and performs a two-dimensional DFT to obtain a sample CSI matrix in the delay angle domain. Furthermore, the UE extracts the real part and the imaginary part from the acquired sample CSI matrix in the delay angle domain. Since the correlation between the real part and the imaginary part of the CSI matrix is similar, the constructed CSI compression encoder may be trained based on the real part or the imaginary part of the sample CSI matrix in the delay angle domain until the training end condition is met to obtain the first target CSI compression encoder. In an embodiment of the present application, only the real part or the imaginary part is used to train the CSI compression encoder, and there is no need to use the real part and the imaginary part at the same time. Compared with the related art in which the real part and the imaginary part of the CSI matrix are taken out separately and stacked, and the whole is used as the input of the neural network model, the size of the input data is reduced, the number of parameters and the amount of calculation of the CSI compression encoder are also reduced, and the training speed is improved.

[0051] The following example illustrates the process of constructing a sample CSI matrix in the delay angle domain:

[0052] The network equipment end samples the uniform linear array (ULA) for downlink transmission, in which the half-wavelength interval is configured as N t = 32 antennas, while the UE is configured with a single antenna and uses N c = 1024 subcarriers, using a specified channel model, such as the COST 2100 MIMO channel model, generates 150,000 spatial-frequency domain sample CSI matrices in a 5.3 GHz indoor picocell scenario and divides them into a training set of 100,000 samples, a validation set of 30,000 samples, and a test set of 20,000 samples. Do a two-dimensional DFT, that is Get the sample CSI matrix H in the delay angle domain a ,in The size is N c ×N t , F a and F b The size is N c ×N c and N t ×N t The DFT matrix of , the superscript H represents the conjugate transpose of the matrix.

[0053] It should be noted that the sample CSI matrix in the space-frequency domain is After DFT processing, the sample CSI matrix H in the delay angle domain is obtained a There are often many rows with extremely small values or zero values. In order to reduce the amount of data processing, H aPerform non-zero principal value preservation to obtain H. The size of H is N cc ×N t , take out the real part and imaginary part of H respectively, and record them as H re and H im , the size is 1×N cc ×N t In the upsampling process, after DFT processing, H a Only the first N cc = 32 non-zero rows, for this H a Retain non-zero primary values, i.e. retain the first N cc non-zero rows, and the matrix after retaining the non-zero principal values is denoted as H.

[0054] For example, the CSI compression encoder may include 4 composite convolutional layers and a fully connected layer, each composite convolutional layer includes a convolutional layer, a batch normalization layer, and an activation function layer. The convolution layer within each composite convolutional layer may include multiple or one convolution kernels, and the convolution kernel sizes may be the same or different, which is not limited in the embodiments of this application. For an introduction to the convolution kernel size, please refer to the relevant content in any embodiment of this application, which will not be repeated here.

[0055] The convolution operation of each composite convolutional layer is defined as:

[0056]

[0057] where y d,i,j is the (d,i,j)th element in the convolution output y, W d,c,h,w is the (d,c,h,w)th element in the convolution kernel weight matrix W, b a is the dth element in the convolution kernel bias b, is the (c,i×s1+h,j×s2+w)th element of the convolution input x after zero padding, s1 and s2 are the convolution steps, denoted as (s1,s2).

[0058] In some implementations, the activation function layer may use a LeakyReLU activation function, where the LeakyReLU activation function is defined as:

[0059]

[0060] It should be noted that the activation function layer may also use other activation functions, which are not limited here.

[0061] Optionally, the full join operation is defined as:

[0062]

[0063] where y iis the i-th element in the fully-connected output, W i,j is the (i, j)-th element in the fully-connected weight matrix, b i is the i-th element in the fully-connected bias, x i is the i-th element in the fully-connected input, the fully-connected output s re and s im is a one-dimensional tensor of size M, let N = N cc ×N t , define the compression ratio γ = M / N, then M < N, γ < 1, to achieve CSI compression.

[0064] To make the matrix sizes consistent, the same zero-padding can be performed on the input of each composite convolutional layer. In some implementations, on both sides of the second dimension i*s1+h of the convolutional input, i.e., the height dimension, add and rows of 0, and on both sides of the third dimension j×s j +w, i.e., the width dimension, add and rows of 0. In the case of the convolutional input 1×N cc ×N t , after convolution, the output sizes of the second and third dimensions are respectively and

[0065] In the embodiments of the present application, before inputting the real part or the imaginary part of the sample CSI matrix in the time-delay angle domain into the first composite convolutional layer, first perform the above-mentioned same zero-padding process, and then input it into the first composite convolutional layer. Further, perform the same zero-padding on the tensor output by the first composite convolutional layer, and then input it into the next composite convolutional layer. The tensors output by each of the remaining composite convolutional layers are all filled with the same zero-padding before being input into the next composite convolutional layer until the N-th composite convolutional layer.

[0066] In some implementations, in order to make the trained first target CSI compression encoder better match the current UE, the UE can obtain an updated CSI matrix in the delay angle domain and, based on the updated CSI matrix, update the first target CSI compression encoder, that is, fine-tune the first target CSI compression encoder to make it more suitable for the current UE. Optionally, the first target CSI compression encoder is updated based on the real part or imaginary part of the updated CSI matrix. In some implementations, when the first target CSI compression encoder is trained based on the real part of the sample CSI matrix, the real part of the updated CSI matrix is used for updating. When the first target CSI compression encoder is trained based on the imaginary part of the sample CSI matrix, the imaginary part of the updated CSI matrix is used for updating. In other implementations, because the real and imaginary parts of the CSI matrix have similar correlations, when the first target CSI compression encoder is trained based on the real part of the sample CSI matrix, the imaginary part of the updated CSI matrix is used for updating, or when the first target CSI compression encoder is trained based on the imaginary part of the sample CSI matrix, the real part of the updated CSI matrix is used for updating.

[0067] It should be noted that the Adam optimization algorithm and end-to-end learning method are used to train the parameters of the first target CSI compression encoder to minimize the cost function. The cost function is described as follows:

[0068]

[0069] Among them, S is the number of samples in the training set, ‖·‖ is the Euclidean norm, and the model parameters mainly include the weight and bias of the fully connected layer, and the weight and bias of the convolution kernel.

[0070] After acquiring the first target CSI compression encoder, the UE may use the first target CSI compression encoder for encoding.

[0071] S302: Encode the target CSI matrix in the delay angle domain based on the first target CSI compression encoder to generate a compressed coding value.

[0072] The specific implementation of step S302 can adopt any implementation method in the embodiments of this application, and will not be repeated here.

[0073] It should be noted that, in the parallel coding method, if the first target CSI compression encoder is an encoder trained based on the real part of the sample CSI matrix, the real part of the target CSI matrix is input into the first target CSI compression encoder for encoding, and the imaginary part of the target CSI matrix is input into the second target CSI compression encoder for encoding. If the first target CSI compression encoder is an encoder trained based on the imaginary part of the sample CSI matrix, the imaginary part of the target CSI matrix is input into the first target CSI compression encoder for encoding, and the real part of the target CSI matrix is input into the second target CSI compression encoder for encoding.

[0074] The CSI encoding method provided in the embodiment of the present application trains the CSI compression coding model through the real part or imaginary part of the sample CSI matrix in the sample delay angle domain, without using the real part and the imaginary part at the same time, thereby reducing the size of the input data, reducing the number of parameters and the amount of calculation of the CSI compression encoder, and improving the training speed. Furthermore, the delay domain convolution step size of the first composite convolution layer in the first target CSI compression encoder deployed on the UE is smaller than the angle domain convolution step size. Since the correlation of the target CSI matrix in the angle domain is strong and the correlation in the delay domain is low, the first composite convolution layer can be used to perform more compression on the angle domain dimension with strong correlation and less compression on the delay domain dimension with low correlation. This not only improves the compression performance of the encoder, but also achieves reasonable compression coding of the target CSI matrix.

[0075] Figure 4 A flow chart of another CSI encoding method provided in an embodiment of the present application. The CSI encoding method is executed by the UE. Figure 4 As shown, the CSI encoding method includes the following steps:

[0076] S401: Obtain a CSI matrix in the space-frequency domain, perform a two-dimensional DFT on the CSI matrix in the space-frequency domain to obtain a target CSI matrix, and retain the main values of non-zero rows.

[0077] The specific implementation of step S401 can adopt any implementation method in the embodiments of this application, and will not be repeated here.

[0078] S402 : Encode the target CSI matrix that has been processed with non-zero row major value retention based on the first target CSI compression encoder to generate a compressed coded value.

[0079] The real and imaginary parts are extracted from the target CSI and compression-encoded based on the first target CSI compression encoder. Optionally, the real and imaginary parts can be compressed and encoded using a serial encoding method or a parallel encoding method. For a detailed description of the serial and parallel encoding methods, please refer to the relevant contents of the embodiments of this application and will not be repeated here.

[0080] S403: Send the compressed code value to the network device.

[0081] In the embodiment of the present application, the real part of the target CSI matrix is compressed and encoded to obtain a low-dimensional matrix s re The imaginary part of the target CSI matrix is compressed and coded to obtain the low-dimensional matrix s im There is a feedback link between the UE and the network device, through which the UE will obtain the low-dimensional matrix s re and s im Transmit to network device.

[0082] In the CSI encoding method provided in the embodiment of the present application, the delay domain convolution step size of the first composite convolution layer in the first target CSI compression encoder deployed on the UE is smaller than the angle domain convolution step size. Since the correlation of the target CSI matrix in the angle domain is strong and the correlation in the delay domain is low, the first composite convolution layer can perform more compression on the angle domain dimension with strong correlation and less compression on the delay domain dimension with low correlation, which not only improves the compression performance of the encoder but also realizes reasonable compression encoding of the target CSI matrix.

[0083] Figure 5 The flowchart of a CSI decoding method provided in an embodiment of the present application is shown in FIG. The CSI decoding method is executed by a network device. Figure 5 As shown, the CSI decoding method includes the following steps:

[0084] S501: Receive a compressed coding value corresponding to a target CSI matrix in a delay angle domain sent by a UE.

[0085] In the embodiment of the present application, the UE performs compression coding on the real and imaginary parts of the target CSI matrix in the delay angle domain based on the first target CSI compression encoder to obtain the compression coding value corresponding to the target CSI matrix. Among them, after the real part of the target CSI matrix is compressed and coded, the low-dimensional matrix s is obtained. re , which is the compressed coding value corresponding to the real part of the target CSI matrix. After the imaginary part of the target CSI matrix is compressed and coded, the low-dimensional matrix s is obtained. im , which is the compression coding value corresponding to the imaginary part of the target CSI matrix. Regarding the process of the UE compressing and coding the real and imaginary parts of the target CSI matrix in the delay angle domain based on the first target CSI compression encoder, please refer to the relevant contents of the above embodiment and will not be repeated here.

[0086] There is a feedback link between the UE and the network device, through which the UE transmits the low-dimensional matrix s re and s imTransmitted to the network device. Correspondingly, the network device receives the low-latitude s feedback from the UE through the feedback link re and s im .

[0087] S502: Decode the compressed coded value based on the first target CSI decoder to generate an estimated value of the target CSI matrix, wherein the first target CSI decoder includes N compound deconvolution layers and one fully connected layer, each compound deconvolution layer includes a delay domain convolution step and an angle domain convolution step, the delay domain convolution step of the Nth compound deconvolution layer among the N compound deconvolution layers is smaller than the angle domain convolution step of the Nth compound deconvolution layer, and N is a positive integer.

[0088] Figure 6 This is a schematic diagram of the structure of the first target CSI decoder provided in the embodiment of the present application. Figure 6 As shown, the first target CSI decoder includes a fully connected layer and N composite deconvolution layers. In the implementation, each of the N composite deconvolution layers includes a deconvolution layer, a batch normalization layer, and an activation function layer.

[0089] It should be noted that the convolution layer within each composite deconvolution layer uses a deconvolution kernel of size f × h × w × c, and a convolution step of (g, p), where f represents the number of deconvolution kernels, h represents the height of the deconvolution kernel, w represents the width of the deconvolution kernel, c represents the number of output channels of the deconvolution kernel, g represents the convolution step in the time domain, and p represents the convolution step in the angle domain. f is an adjustable parameter that can be adjusted according to usage requirements. f, h, w, c, g, and p are all positive integers.

[0090] Optionally, the Nth composite deconvolution layer uses a deconvolution layer with a deconvolution kernel size of f×1×m×n and a convolution step size of (k, l), where f corresponds to the number of deconvolution kernels, 1 corresponds to the height of the deconvolution kernel, m corresponds to the width of the deconvolution kernel, n corresponds to the number of input channels of the deconvolution kernel, k corresponds to the convolution step size in the time domain, and l corresponds to the convolution step size in the angle domain. f is an adjustable parameter that can be adjusted according to usage requirements, and f, m, n, k, and l are all positive integers.

[0091] In this application example, the CSI compression encoder deployed in the UE limits the first composite deconvolution layer to satisfying the requirement that the delay domain convolution step size k is smaller than the angle domain convolution step size l. Accordingly, in order to accurately decode the target CSI matrix, the Nth composite deconvolution layer in the first target CSI decoder needs to satisfy the requirement that the delay domain convolution step size k is smaller than the angle domain convolution step size l.

[0092] In the embodiment of the present application, the Nth composite deconvolution layer also needs to meet the following conditions:

[0093] The delay domain convolution step size of the Nth composite deconvolution layer is smaller than the width of the deconvolution kernel in the Nth composite deconvolution layer. That is, the delay domain convolution step size k in the Nth composite deconvolution layer is smaller than the width m of the deconvolution kernel.

[0094] The angle domain convolution step size of the Nth composite deconvolution layer is smaller than the number of input channels of the deconvolution kernel in the Nth composite deconvolution layer. In other words, the angle domain convolution step size l in the Nth composite deconvolution layer is smaller than the width n of the deconvolution kernel.

[0095] The network device is deployed with a first target CSI decoder, which inputs the received compressed coded value into the first target CSI decoder for decoding to obtain an estimated value of the target CSI matrix. In other words, the network device converts the low-dimensional matrix s corresponding to the real part of the target CSI matrix into re and / or the low-dimensional matrix s corresponding to the imaginary part of the target CSI matrix im , input into the first target CSI decoder for decoding, and obtain the real and imaginary part estimates of the delay angle domain CSI matrix, and

[0096] Optionally, the network device converts the low-dimensional matrix s re and / or the low-dimensional matrix s corresponding to the imaginary part im , input to the fully connected layer in the first target CSI decoder, the fully connected layer outputs a vector, and further, the vector is reconstructed to generate the first tensor. In some implementations, the low-dimensional matrix s corresponding to the real part of the target CSI matrix is converted to re , or the low-dimensional matrix s corresponding to the imaginary part im , respectively input into the fully connected layer, or, the corresponding low-dimensional matrix s of the target CSI matrix re and the low-dimensional matrix s corresponding to the imaginary part im Input into the fully connected layer in sequence.

[0097] After obtaining the first tensor, the first tensor is input into the first composite deconvolution layer to obtain the second tensor. Furthermore, the tensor output by the previous composite deconvolution layer is used as the input of the next composite deconvolution layer, until the Nth composite deconvolution layer, where the tensor output by the Nth composite deconvolution layer is the estimated value of the target CSI. It should be noted that the size of the tensor output by the previous composite deconvolution layer may change under the influence of the convolution step size of the next composite deconvolution layer after the convolution process. re , and s im Input the fully connected layer and output a 1×(N cc ×(N t / 2)) and reconstruct a vector of size 1×N cc ×(Nt / 2). Taking the first composite deconvolution layer as an example, the size of the first tensor is 1×N cc ×(N t / 2) Input it to the second composite deconvolution layer. After deconvolution, the size of the second tensor output can be f×N cc ×(N t / 2).

[0098] Optionally, in a scenario where the network device deploys only the first target CSI decoder, since the real and imaginary parts of the CSI matrix have similar correlations, the first target CSI decoder can be used to decode the compressed coded values of the real and imaginary parts of the target CSI matrix. In some implementations, the network device sequentially inputs the compressed coded values of the target CSI matrix and the imaginary parts into the first target CSI decoder for decoding. In other words, the network device serially inputs the real and imaginary parts of the target CSI matrix into the first target CSI decoder for decoding. In the serial decoding method, since only one target CSI decoder is deployed, the serial encoding method occupies fewer resources.

[0099] Since the correlation between the real and imaginary parts of the CSI matrix is similar, the parameters of the decoders used when decoding the real and imaginary parts are often similar or identical. Therefore, optionally, based on the model parameters of the first target CSI decoder, a second target CSI decoder is constructed, and the compressed coding value of the real part and the compressed coding value of the imaginary part of the target CSI matrix are respectively input into their respective corresponding target CSI decoders for decoding. In other words, the UE deploys two target CSI decoders at the same time, and can input the compressed coding value of the real part of the target CSI matrix into one of the target CSI decoders, and in parallel input the compressed coding value of the imaginary part of the target CSI matrix into the other target CSI decoder. In the embodiment of the present application, in the parallel decoding method, since two target CSI decoders are deployed at the same time, the real and imaginary parts of the target CSI matrix can be decoded synchronously, making the decoding speed of the parallel decoding method faster.

[0100] It can be understood that if the compressed coded value of the real part corresponds to the first target CSI decoder, then the compressed coded value of the imaginary part corresponds to the second target CSI decoder, or, if the compressed coded value of the real part corresponds to the second target CSI decoder, then the compressed coded value of the imaginary part corresponds to the first target CSI decoder. Optionally, if the first target CSI decoder is trained using the compressed coded value of the real part of the sample CSI matrix, then the compressed coded value of the real part of the target CSI matrix corresponds to the first target CSI decoder, and the compressed coded value of the imaginary part of the target CSI matrix corresponds to the second target CSI decoder. If the first target CSI decoder is trained using the compressed coded value of the imaginary part of the sample CSI matrix, then the compressed coded value of the imaginary part of the target CSI matrix corresponds to the first target CSI decoder, and the compressed coded value of the real part of the target CSI matrix corresponds to the second target CSI decoder.

[0101] In the CSI decoding method provided in the embodiment of the present application, the delay domain convolution step size of the first composite convolution layer in the first target CSI compression encoder deployed on the UE is smaller than the angle domain convolution step size. Since the correlation of the target CSI matrix in the angle domain is strong and the correlation in the delay domain is low, the first composite convolution layer can be used to perform more compression on the angle domain dimension with strong correlation, and less compression on the delay domain dimension with low correlation. Accordingly, the delay domain convolution step size of the Nth composite deconvolution layer in the first target CSI decoder deployed on the network device is smaller than the angle domain convolution step size, so that the Nth composite deconvolution layer can be used to perform more recovery on the angle domain dimension with strong correlation, and less recovery on the delay domain dimension with low correlation, thereby achieving accurate decoding of the compressed coding value of the target CSI matrix.

[0102] Figure 7 A flowchart of another CSI decoding method provided in an embodiment of the present application. The CSI decoding method is executed by a network device. Figure 7 As shown, the CSI decoding method includes the following steps:

[0103] S701: Receive a sample compression code value sent by a UE, where the sample compression code value is generated by the UE encoding a real part or an imaginary part of a sample CSI matrix in a delay angle domain based on a CSI compression encoder.

[0104] The UE uses a CSI compression encoder to encode the real or imaginary part of the sample CSI matrix in the delay angle domain to generate a sample compression value corresponding to the real part or a sample compression encoding value corresponding to the imaginary part. For a detailed introduction to the CSI compression encoder, please refer to the relevant content in the above embodiments and will not be repeated here. The process of the UE performing compression encoding on the real or imaginary part of the sample CSI matrix in the delay angle domain based on the CSI compression encoder can be referred to the relevant content in the above embodiments and will not be repeated here.

[0105] S702: Train a CSI decoder based on the sample compression coding value to obtain a first target CSI decoder.

[0106] In order to realize the decoding of CSI matrix coding, it is necessary to pre-train a first target CSI decoder, and at the beginning of training, randomly initialize the parameters of each layer. In an embodiment of the present application, an initial CSI decoder is pre-constructed, and the CSI decoder includes a fully connected layer and N compound deconvolution layers, wherein each compound deconvolution layer includes a delay domain convolution step and an angle domain convolution step, and the delay domain convolution step of the Nth compound deconvolution layer among the N compound deconvolution layers is smaller than the angle domain convolution step of the Nth compound deconvolution layer, and N is a positive integer.

[0107] Optionally, the Nth composite deconvolution layer also needs to meet the following conditions:

[0108] The delay domain convolution step size of the Nth composite deconvolution layer is smaller than the width of the deconvolution kernel in the Nth composite deconvolution layer;

[0109] The angle domain convolution step size of the Nth composite deconvolution layer is smaller than the number of input channels of the deconvolution kernel in the Nth composite deconvolution layer.

[0110] In the embodiment of the present application, since the correlation between the real part and the imaginary part of the CSI matrix is similar, the constructed CSI decoder can be trained based on the compressed coding value of the real part or the compressed coding value of the imaginary part of the sample CSI matrix in the delay angle domain until the training end condition is met to obtain the first target CSI decoder. Optionally, the training end condition can be the output of the network device CSI decoder. With H re The gap is the smallest, or the number of training times reaches the preset number.

[0111] In the embodiment of the present application, only the compressed coded value of the real part or the compressed coded value of the imaginary part is used to train the CSI decoder, and there is no need to use the compressed coded value of the real part and the imaginary part at the same time. Compared with the related art of taking out the compressed coded value of the real part and the compressed coded value of the imaginary part of the CSI matrix and stacking them separately, and using the whole as the input of the neural network model, the size of the input data is reduced, the number of parameters and the amount of calculation of the CSI compression encoder are also reduced, and the training speed is improved.

[0112] For example, the CSI decoder may include a fully connected layer and four compound deconvolution layers, each compound deconvolution layer including a deconvolution layer, a batch normalization layer, and an activation function layer. The deconvolution layer within each compound deconvolution layer may include multiple or one deconvolution kernels, and the deconvolution kernel sizes may be the same or different, which is not limited in the embodiments of the present application. For an introduction to the deconvolution kernel size, please refer to the relevant contents in any embodiment of the present application, which will not be repeated here.

[0113] The convolution operation of each composite deconvolution layer is defined as:

[0114]

[0115] in is the deconvolution output y D The (d,i,j)th element in is the deconvolution kernel weight matrix W D The (d,c,h,w)th element in is the deconvolution kernel bias b D The dth element in Input x for deconvolution D After zero-padding preprocessing and zero-padding, the (c,i+h,j+w)th element, and is the deconvolution step length, denoted as The superscript D only indicates relevance to deconvolution.

[0116] In some implementations, the first N-1 of the N composite deconvolution layers use the LeakyReLU activation function, and the Nth composite deconvolution layer uses the logistic function as the activation function. It should be noted that the activation function layer can also use other activation functions, which are not limited here.

[0117] To make the matrix size consistent, the input of each composite deconvolution layer can be padded with zeros. In some implementations, the second dimension of the deconvolution input, i+h, is inserted between each element of the height dimension. 0. The third dimension, j+w, is added between each element of the width dimension 0. After the deconvolution input is preprocessed with zero padding, the number of zero paddings is the same as the convolution zero padding with a convolution stride of (1, 1).

[0118] In the embodiment of the present application, the output vector of the fully connected layer is reconstructed and input into the first composite deconvolution layer, and then undergoes the same zero-padding process as described above. The tensor output by the first composite deconvolution layer is further padded with the same zeros before being input into the next composite deconvolution layer. The tensors output by each remaining composite deconvolution layer are padded with the same zeros before being input into the next composite deconvolution layer, and so on until the Nth composite deconvolution layer.

[0119] In some implementations, in order to make the trained first target CSI decoder better match the current network device, the network device can receive an updated compressed coding value of the updated CSI matrix in the delay angle domain through a feedback link, and based on the updated compressed coding value, update the first target CSI decoder, that is, fine-tune the first target CSI decoder so that the first target CSI decoder is more suitable for the current network device.

[0120] Optionally, the network device updates the first target CSI decoder based on the updated compressed coded value corresponding to the real part or the updated compressed coded value corresponding to the imaginary part of the updated CSI matrix. In some implementations, when the first target CSI decoder is trained based on the sample compressed coded value of the real part of the sample CSI matrix, the first target CSI decoder is updated using the updated compressed coded value of the real part of the updated CSI matrix. When the first target CSI decoder is trained based on the imaginary part of the sample CSI matrix, the first target CSI decoder is updated using the updated compressed coded value of the imaginary part of the updated CSI matrix. In other implementations, because the real and imaginary parts of the CSI matrix have similar correlations, when the first target CSI decoder is trained based on the sample compressed coded value of the real part of the sample CSI matrix, the first target CSI decoder is updated using the updated compressed coded value of the imaginary part of the updated CSI matrix. Alternatively, when the first target CSI decoder is trained based on the sample compressed coded value of the imaginary part of the sample CSI matrix, the first target CSI decoder is updated using the compressed coded value of the real part of the updated CSI matrix.

[0121] It should be noted that the Adam optimization algorithm and end-to-end learning method are used to train the parameters of the first target CSI decoder to minimize the cost function. The cost function is described as follows:

[0122]

[0123] Among them, S is the number of samples in the training set, ‖·‖ is the Euclidean norm, and the model parameters mainly include the weight and bias of the fully connected layer, and the weight and bias of the deconvolution kernel.

[0124] After acquiring the first target CSI decoder, the UE may use the first target CSI decoder for decoding.

[0125] S703: Receive a compressed coding value corresponding to a target CSI matrix in the delay angle domain sent by the UE.

[0126] S704: Decode the compressed coded value based on the first target CSI decoder to generate an estimated value of the target CSI matrix.

[0127] The specific implementation of steps S703 to S704 may adopt any implementation method in the embodiments of the present application, and will not be described in detail here.

[0128] Optionally, the CSI decoding method provided in the embodiment of the present application may further include:

[0129] S705 , performing inverse DFT on the estimated value of the target CSI matrix to obtain a space-frequency domain CSI matrix.

[0130] The network device converts the estimated value of the real part of the target CSI matrix output by the target CSI decoder into and an estimate of the imaginary part Combined into a complex matrix Through the two-dimensional inverse DFT, the reconstruction value of the original spatial frequency domain CSI matrix is obtained

[0131] In the CSI decoding method provided in the embodiment of the present application, the delay domain convolution step size of the first composite convolution layer in the first target CSI compression encoder deployed on the UE is smaller than the angle domain convolution step size. Since the correlation of the target CSI matrix in the angle domain is strong and the correlation in the delay domain is low, the first composite convolution layer can be used to perform more compression on the angle domain dimension with strong correlation, and less compression on the delay domain dimension with low correlation. Accordingly, the delay domain convolution step size of the Nth composite deconvolution layer in the first target CSI decoder deployed on the network device is smaller than the angle domain convolution step size, so that the Nth composite deconvolution layer can be used to perform more recovery on the angle domain dimension with strong correlation, and less recovery on the delay domain dimension with low correlation, thereby achieving accurate decoding of the compressed coding value of the target CSI matrix.

[0132] It should be noted that the number dimension f of the convolution kernel and deconvolution kernel in the above embodiment can be selected according to the actual application requirements. If the feedback accuracy is required to be high and the parameter quantity is required to be low, f=32 can be selected. If the parameter quantity is required to be as small as possible and the feedback accuracy can be appropriately reduced, f=16 can be selected.

[0133] Figure 8FIG. 1 is a flow chart of a CSI transmission method provided in an embodiment of the present application. Figure 8 As shown, the UE compresses and encodes the target CSI matrix using a serial encoding method, and the network device decodes the target CSI matrix using a serial decoding method. The ENet model includes a first target CSI compression encoder and a first target CSI decoder.

[0134] UE converts the CSI matrix H in the space-frequency domain f Perform two-dimensional DFT processing to obtain the target CSI matrix H in the delay angle domain a , for H a Retain the non-zero principal value to obtain H, and take out the real and imaginary parts of H respectively, and record them as H re and H im .

[0135] Furthermore, the UE will re and H im Serial input to the ENet model, the first target CSI compression encoder of the ENet model sequentially re and H im Serial encoding, generating H re The compressed encoding value s re , and H im The compressed encoding value s im .

[0136] UE will s re and s im It is transmitted to the network device through the feedback link and is sequentially s by the first target CSI decoder of the ENet model. re and s im Serial decoding and reconstruction to generate H re Estimated value of and H im Estimated value of Furthermore, H re Estimated value of and H im Estimated value of Combined into a complex matrix Complex matrix Perform zero value recovery to generate the reconstruction matrix of the delay angle domain Furthermore, the reconstruction matrix Perform two-dimensional inverse DFT processing to obtain the reconstructed value of the original spatial frequency domain CSI matrix

[0137] Figure 9 A flow chart of another CSI transmission method provided in an embodiment of the present application. Figure 9As shown, the UE uses parallel coding to compress and encode the target CSI matrix, and the network device uses parallel decoding to decode the target CSI matrix. ENet model 1 includes a first target CSI compression encoder and a first target CSI decoder, and ENet model 2 includes a second target CSI compression encoder and a second target CSI decoder. The parameters of the first target CSI compression encoder and the second target CSI compression encoder are the same, and the parameters of the first target CSI decoder and the second target CSI decoder are the same.

[0138] UE converts the CSI matrix H in the space-frequency domain f Perform two-dimensional DFT processing to obtain the target CSI matrix H in the delay angle domain a , for H a Retain the non-zero principal value to obtain H, and take out the real and imaginary parts of H respectively, and record them as H re and H im .

[0139] Furthermore, the UE will re Input into Eet model 1, and in parallel H im Input to ENet model 2. The first target CSI compression encoder of ENet model 1 is H re Encode, in parallel, the second target CSI compression encoder in ENet model 2 is H im Encode and generate H re The compressed encoding value s re , and H im The compressed encoding value s im .

[0140] UE will s re and s im The first target CSI decoder in ENet model 1 transmits s in parallel to the network device through the feedback link. re Decode and reconstruct H re Estimated value of And the second target CSI decoder in ENet model 2 is used to decode s im Decode and reconstruct H im Estimated value of Furthermore, H re Estimated value of and H im Estimated value of Combined into a complex matrix Complex matrix Perform zero value recovery to generate the reconstruction matrix of the delay angle domain Furthermore, the reconstruction matrix Perform two-dimensional inverse DFT processing to obtain the reconstructed value of the original spatial frequency domain CSI matrix

[0141] Figure 10 This is a schematic diagram of an application of a CSI transmission method provided in an embodiment of the present application. Figure 10 As shown in the figure, the UE is deployed with a target CSI compression encoder, and the network equipment is deployed with a target CSI decoder. The target CSI compression encoder includes four composite convolutional layers and one fully connected layer. The target CSI decoder includes one fully connected layer and four composite deconvolutional layers.

[0142] For example, the convolution layer of the first composite convolution layer in the target CSI compression encoder uses a convolution kernel of size f×1×3×5 and a convolution step of (1, 2), that is, m=3, k=5, k=1, l=2. The convolution layers of the second and third composite convolution layers both use a convolution kernel of size f×f×3×3 and a convolution step of (1, 1). The convolution layer of the fourth composite convolution layer uses a convolution kernel of size 1×f×3×3 and a convolution step of (1, 1). The UE extracts the real part H from the CSI matrix H in the delay angle domain. re and the imaginary part H im , H re and / or H im Input CSI compression encoder, where H re and H im The size is 1×N cc ×N t The convolution layer of the first composite convolution layer of the CSI compression encoder is H re or H im After convolution processing, the output tensor size is f×N cc ×(N t / 2). The output tensor of the first composite convolutional layer is of size f×N cc ×(N t / 2) Input into the second composite convolution layer. After convolution processing by the convolution layer of the second composite convolution layer, the output tensor size is f×N cc ×(N t / 2). Then, the output tensor of the second composite convolutional layer is of size f×N. cc ×(N t / 2) Input into the third composite convolution layer. After convolution processing by the convolution layer of the third composite convolution layer, the output tensor size is f×N cc ×(N t / 2). Further, the output tensor size of the third composite convolutional layer is set to f×N cc ×(N t / 2) Input into the fourth composite convolution layer, after convolution processing by the convolution layer of the fourth composite convolution layer, the output tensor size is 1×N cc ×(N t / 2). Reshape the output tensor into a 1×(1×N cc ×(N t / 2)) is input into the fully connected layer for full connection operation, and the output s is obtained. re or s im is a one-dimensional tensor of size M.

[0143] UE will s re or s im It is sent to the network device through the feedback link, and the network device will s re or s im Input to the CSI decoder.

[0144] The first composite deconvolution layer in the CSI decoder uses a deconvolution kernel of size f×1×3×3. The second and third composite deconvolution layers use a deconvolution kernel of size f×f×3×3, with a deconvolution step of (1,1). The fourth composite deconvolution layer uses a deconvolution kernel of size f×1×3×5, with a deconvolution step of (1,2). The network device receives the s re or s im Input the fully connected layer for full connection processing, and the output is 1×(N cc ×(N t / 2)) and reconstruct a vector of size 1×N cc ×(N t / 2) as the input of the first composite deconvolution layer. The deconvolution layer of the first composite deconvolution layer is 1×N cc ×(N t / 2) after deconvolution, the output tensor size is f×N cc ×(N t / 2), the output tensor size of the first composite deconvolution layer is f×N cc ×(N t / 2) Input into the second composite deconvolution layer. After deconvolution processing by the convolution layer of the second composite deconvolution layer, the output tensor size is f×N cc ×(N t / 2). Then the output tensor of the second composite deconvolution layer is f×N cc ×(N t / 2) Input into the third composite deconvolution layer. After deconvolution processing by the convolution layer of the third composite deconvolution layer, the output tensor size is f×N cc ×(Nt / 2). The output tensor of the third composite deconvolution layer is of size f×N cc ×(N t / 2) Input the fourth composite deconvolution layer. After deconvolution processing by the convolution layer of the fourth composite deconvolution layer, the output tensor size is 1×N cc ×N t , that is, the real and imaginary part estimates of the original delay angle domain CSI matrix, or

[0145] Optionally, the network device converts the estimated value of the real part of the output of the CSI decoder and an estimate of the imaginary part Combined into a complex matrix Through the two-dimensional inverse DFT, the reconstruction value of the original spatial frequency domain CSI matrix is obtained

[0146] Corresponding to the CSI encoding methods provided in the above-mentioned embodiments, the present application also provides a CSI encoding device. The implementation of the CSI encoding method is also applicable to the CSI encoding device provided in this embodiment and will not be described in detail in this embodiment.

[0147] Figure 11 This is a schematic diagram of the structure of a CSI encoding device provided in an embodiment of the present application. Figure 11 As shown, the CSI encoding device 100 includes: an encoding module 11.

[0148] Among them, the encoding module 11 is used to encode the target CSI matrix in the delay angle domain based on the first target CSI compression encoder to generate a compressed coding value, wherein the first target CSI compression encoder includes N compound convolution layers and a fully connected layer, each compound convolution layer includes a delay domain convolution step and an angle domain convolution step, the delay domain convolution step of the first compound convolution layer among the N compound convolution layers is smaller than the angle domain convolution step of the first compound convolution layer, and N is a positive integer.

[0149] The first target CSI compression encoder includes N composite convolutional layers and a fully connected layer. In the implementation, each of the N composite convolutional layers includes a convolutional layer, a batch normalization layer, and an activation function layer.

[0150] Because the target CSI matrix in the delay angle domain has different correlations in the delay and angle domains—that is, the correlation is stronger in the angle domain and lower in the delay domain—in order to properly compress and encode the target CSI matrix, in this application example, the first composite convolution layer in the first target CSI compression encoder is required to ensure that the delay domain convolution step size k is smaller than the angle domain convolution step size l. This restriction allows for greater compression of the highly correlated angle domain dimension and less compression of the less correlated delay domain dimension, resulting in better compression performance for the encoder.

[0151] In this embodiment of the present application, the first composite convolutional layer also needs to meet the following conditions:

[0152] The delay domain convolution step size of the first composite convolution layer is smaller than the width of the convolution kernel in the first composite convolution layer. That is, the delay domain convolution step size k in the first composite convolution layer is smaller than the width m of the convolution kernel.

[0153] The angular domain convolution step size of the first composite convolution layer is smaller than the number of input channels of the convolution kernel in the first composite convolution layer. In other words, the angular domain convolution step size l in the first composite convolution layer is smaller than the width n of the convolution kernel.

[0154] UE converts the CSI matrix H in the space-frequency domain f Perform a two-dimensional discrete Fourier transform (DFT) to obtain a target CSI matrix in the delay angle domain. Further, the target CSI matrix is input into a first target CSI compression encoder, and the first target CSI compression encoder compresses and encodes the target CSI matrix to obtain a compressed coding value. Optionally, the real part and / or imaginary part of the target CSI matrix is input into the first composite convolution layer in the first target CSI compression encoder to obtain a tensor output by the first composite convolution layer. In some implementations, the real part or imaginary part of the target CSI matrix is input into the first composite convolution layer, respectively, or the real part and imaginary part of the target CSI matrix are input into the first composite convolution layer in sequence.

[0155] Furthermore, the tensor output by the previous composite convolution layer is used as the input of the next composite convolution layer until the Nth composite convolution layer. Finally, the tensor output by the Nth composite convolution layer is reconstructed, and the reconstructed tensor is input to the fully connected layer, which performs a full connection operation to generate a compressed coding value of the target CSI matrix.

[0156] Optionally, in a scenario where the UE deploys only the first target CSI compression encoder, since the real and imaginary parts of the CSI matrix have similar correlations, the first target CSI compression encoder can be used to compress and encode the real and imaginary parts of the target CSI matrix. In some implementations, the UE can extract the real and imaginary parts of the target CSI matrix and input the real and imaginary parts of the target CSI matrix into the first target CSI compression encoder in sequence for encoding. In other words, the UE serially inputs the real and imaginary parts of the target CSI matrix into the first target CSI compression encoder for encoding. In the serial encoding method, since only one target CSI compression encoder is deployed, the serial encoding method occupies fewer resources.

[0157] Since the correlation between the real part and the imaginary part of the CSI matrix is similar, the parameters of the compression encoders used when encoding the real part and the imaginary part are often similar or the same. Therefore, optionally, based on the model parameters of the first target CSI compression encoder, a second target CSI compression encoder is constructed to extract the real part and the imaginary part of the target CSI matrix, and the real part and the imaginary part of the target CSI matrix are respectively input into the corresponding target CSI compression encoders for encoding. In other words, the UE deploys two target CSI compression encoders at the same time, and can input the real part of the target CSI matrix into one of the target CSI compression encoders, and input the imaginary part of the target CSI matrix into the other target CSI compression encoder in parallel. In the embodiment of the present application, in the parallel coding method, since two target CSI compression encoders are deployed at the same time, the real part and the imaginary part of the target CSI matrix can be encoded synchronously, so that the encoding speed of the parallel coding method is faster.

[0158] It can be understood that if the real part corresponds to the first target CSI compression encoder, the imaginary part corresponds to the second target CSI compression encoder, or, if the real part corresponds to the second target CSI compression encoder, the imaginary part corresponds to the first target CSI compression encoder. Optionally, if the real part of the sample CSI matrix is used to train the first target CSI compression encoder, then the real part of the target CSI matrix corresponds to the first target CSI compression encoder, and the imaginary part of the target CSI matrix corresponds to the first target CSI compression encoder. If the imaginary part of the sample CSI matrix is used to train the first target CSI compression encoder, then the imaginary part of the target CSI matrix corresponds to the first target CSI compression encoder, and the real part of the target CSI matrix corresponds to the second target CSI compression encoder.

[0159] In the CSI encoding device provided in the embodiment of the present application, the delay domain convolution step size of the first composite convolution layer in the first target CSI compression encoder deployed on the UE is smaller than the angle domain convolution step size. Since the target CSI matrix has a strong correlation in the angle domain and a low correlation in the delay domain, the first composite convolution layer can perform more compression on the angle domain dimension with strong correlation and less compression on the delay domain dimension with low correlation, which not only improves the compression performance of the encoder but also realizes reasonable compression encoding of the target CSI matrix.

[0160] Figure 12 This is a schematic diagram of the structure of a CSI encoding device provided in an embodiment of the present application. Figure 12 As shown, the CSI encoding device 200 includes: an encoding module 21, a training module 22, an updating module 23, a conversion module 24 and a sending module 25.

[0161] Among them, the encoding module 21 is used to encode the target CSI matrix in the delay angle domain based on the first target CSI compression encoder to generate a compressed coding value, wherein the first target CSI compression encoder includes N compound convolution layers and a fully connected layer, each compound convolution layer includes a delay domain convolution step and an angle domain convolution step, the delay domain convolution step of the first compound convolution layer among the N compound convolution layers is smaller than the angle domain convolution step of the first compound convolution layer, and N is a positive integer.

[0162] The training module 22 is configured to train the CSI compression encoder based on the real part or the imaginary part of the sample CSI matrix in the delay angle domain to obtain the first target CSI compression encoder.

[0163] The updating module 23 is configured to obtain an updated CSI matrix in the delay angle domain, and update the first target CSI compression encoder based on the updated CSI matrix.

[0164] Optionally, the encoding module 21 is further configured to:

[0165] constructing a second target CSI compression encoder based on the model parameters of the first target CSI compression encoder;

[0166] Extracting the real and imaginary parts of the target CSI matrix;

[0167] The real part and the imaginary part of the target CSI matrix are respectively input into corresponding target CSI compression encoders for encoding.

[0168] Optionally, the encoding unit 213 is further configured to:

[0169] If the first target CSI compression encoder is an encoder trained based on the real part of the sample CSI matrix, the real part of the target CSI matrix is input into the first target CSI compression encoder for encoding, and the imaginary part of the target CSI matrix is input into the second target CSI compression encoder for encoding; or

[0170] If the first target CSI compression encoder is an encoder trained based on the imaginary part of the sample CSI matrix, the imaginary part of the target CSI matrix is input into the first target CSI compression encoder for encoding, and the real part of the target CSI matrix is input into the second target CSI compression encoder for encoding.

[0171] Optionally, the encoding module 21 is further configured to:

[0172] The real part and the imaginary part of the target CSI matrix are extracted, and the real part and the imaginary part are sequentially input into the first target CSI compression encoder for encoding.

[0173] Optionally, the encoding process performed by the first target CSI compression encoder includes:

[0174] Inputting the real part and / or imaginary part of the target CSI matrix into the first composite convolutional layer to obtain a tensor output by the first composite convolutional layer;

[0175] Using the tensor output by the previous composite convolutional layer as the input of the next composite convolutional layer, until the Nth composite convolutional layer;

[0176] The tensor output by the Nth composite convolutional layer is reconstructed, and the reconstructed vector is input into a fully connected layer to generate the compressed coded value.

[0177] Optionally, each of the N composite convolutional layers includes a convolution layer, a batch normalization layer, and an activation function layer.

[0178] Optionally, the encoding module 21 is further configured to perform the same zero-padding processing on the input of each composite convolutional layer.

[0179] Optionally, the conversion module 24 is configured to obtain a CSI matrix in the space-frequency domain, perform a two-dimensional discrete Fourier transform (DFT) on the CSI matrix to obtain the target CSI matrix, and retain non-zero row principal values.

[0180] Optionally, the sending module 25 is configured to send the compressed encoding value to a network device.

[0181] The CSI encoding device provided in the embodiment of the present application trains the CSI compression coding model through the real part or imaginary part of the sample CSI matrix in the sample delay angle domain, without using both the real and imaginary parts simultaneously, thereby reducing the size of the input data, reducing the number of parameters and the amount of computation of the CSI compression encoder, and improving the training speed. Furthermore, the delay domain convolution step size of the first composite convolution layer in the first target CSI compression encoder of the UE is smaller than the angle domain convolution step size. Since the correlation of the target CSI matrix in the angle domain is strong and the correlation in the delay domain is low, the first composite convolution layer can perform more compression on the angle domain dimension with strong correlation and perform less compression on the delay domain dimension with low correlation, which not only improves the compression performance of the encoder, but also achieves reasonable compression coding of the target CSI matrix.

[0182] Corresponding to the CSI decoding methods provided in the above-mentioned embodiments, the present application also provides a CSI decoding device. The implementation of the CSI decoding method is also applicable to the CSI decoding device provided in this embodiment and will not be described in detail in this embodiment.

[0183] Figure 13 This is a schematic diagram of the structure of a CSI decoding device provided in an embodiment of the present application. Figure 13 As shown, the CSI decoding device 300 includes: a receiving module 31 and a decoding module 32.

[0184] The receiving module 31 is configured to receive a compressed coding value corresponding to a target CSI matrix in a delay angle domain sent by a UE;

[0185] A decoding module 32 is used to decode the compressed coding value based on a first target CSI decoder to generate an estimated value of the target CSI matrix, wherein the first target CSI decoder includes N compound deconvolution layers and one fully connected layer, each compound deconvolution layer includes a delay domain convolution step and an angle domain convolution step, the delay domain convolution step of the Nth compound deconvolution layer among the N compound deconvolution layers is smaller than the angle domain convolution step of the Nth compound deconvolution layer, and N is a positive integer.

[0186] In the embodiment of the present application, the UE performs compression coding on the real and imaginary parts of the target CSI matrix in the delay angle domain based on the first target CSI compression encoder to obtain the compression coding value corresponding to the target CSI matrix. Among them, after the real part of the target CSI matrix is compressed and coded, the low-dimensional matrix s is obtained. re , which is the compressed coding value corresponding to the real part of the target CSI matrix. After the imaginary part of the target CSI matrix is compressed and coded, the low-dimensional matrix s is obtained. im, which is the compression coding value corresponding to the imaginary part of the target CSI matrix. Regarding the process of the UE compressing and coding the real and imaginary parts of the target CSI matrix in the delay angle domain based on the first target CSI compression encoder, please refer to the relevant contents of the above embodiment and will not be repeated here.

[0187] There is a feedback link between the UE and the network device, through which the UE transmits the low-dimensional matrix s re and s im Transmitted to the network device. Correspondingly, the network device receives the low-latitude s feedback from the UE through the feedback link re and s im .

[0188] The first target CSI decoder includes a fully connected layer and N composite deconvolution layers. In the implementation, each of the N composite deconvolution layers includes a deconvolution layer, a batch normalization layer, and an activation function layer.

[0189] In this application example, the CSI compression encoder deployed in the UE limits the first composite deconvolution layer to satisfying the requirement that the delay domain convolution step size k is smaller than the angle domain convolution step size l. Accordingly, in order to accurately decode the target CSI matrix, the Nth composite deconvolution layer in the first target CSI decoder needs to satisfy the requirement that the delay domain convolution step size k is smaller than the angle domain convolution step size l.

[0190] In the embodiment of the present application, the Nth composite deconvolution layer also needs to meet the following conditions:

[0191] The delay domain convolution step size of the Nth composite deconvolution layer is smaller than the width of the deconvolution kernel in the Nth composite deconvolution layer. That is, the delay domain convolution step size k in the Nth composite deconvolution layer is smaller than the width m of the deconvolution kernel.

[0192] The angle domain convolution step size of the Nth composite deconvolution layer is smaller than the number of input channels of the deconvolution kernel in the Nth composite deconvolution layer. In other words, the angle domain convolution step size l in the Nth composite deconvolution layer is smaller than the width n of the deconvolution kernel.

[0193] The network device is deployed with a first target CSI decoder, which inputs the received compressed coded value into the first target CSI decoder for decoding to obtain an estimated value of the target CSI matrix. In other words, the network device converts the low-dimensional matrix s corresponding to the real part of the target CSI matrix into re and / or the low-dimensional matrix s corresponding to the imaginary part of the target CSI matrix im , input into the first target CSI decoder for decoding, and obtain the real and imaginary part estimates of the delay angle domain CSI matrix, and

[0194] Optionally, the network device converts the low-dimensional matrix s re and / or the low-dimensional matrix s corresponding to the imaginary part im , input to the fully connected layer in the first target CSI decoder, the fully connected layer outputs a vector, and further, the vector is reconstructed to generate the first tensor. In some implementations, the low-dimensional matrix s corresponding to the real part of the target CSI matrix is converted to re , or the low-dimensional matrix s corresponding to the imaginary part im , respectively input into the fully connected layer, or, the corresponding low-dimensional matrix s of the target CSI matrix re and the low-dimensional matrix s corresponding to the imaginary part im Input into the fully connected layer in sequence.

[0195] After obtaining the first tensor, the first tensor is input into the first composite deconvolution layer to obtain the second tensor. Furthermore, the tensor output by the previous composite deconvolution layer is used as the input to the next composite deconvolution layer, up to the Nth composite deconvolution layer, where the tensor output by the Nth composite deconvolution layer is the estimated value of the target CSI. It should be noted that the size of the tensor output by the previous composite deconvolution layer may change due to the convolution step size of the next composite deconvolution layer.

[0196] Optionally, in a scenario where the network device deploys only the first target CSI decoder, since the real and imaginary parts of the CSI matrix have similar correlations, the first target CSI decoder can be used to decode the compressed coded values of the real and imaginary parts of the target CSI matrix. In some implementations, the network device sequentially inputs the compressed coded values of the target CSI matrix and the imaginary parts into the first target CSI decoder for decoding. In other words, the network device serially inputs the real and imaginary parts of the target CSI matrix into the first target CSI decoder for decoding. In the serial decoding method, since only one target CSI decoder is deployed, the serial encoding method occupies fewer resources.

[0197] Since the correlation between the real and imaginary parts of the CSI matrix is similar, the parameters of the decoders used when decoding the real and imaginary parts are often similar or identical. Therefore, optionally, based on the model parameters of the first target CSI decoder, a second target CSI decoder is constructed, and the compressed coding value of the real part and the compressed coding value of the imaginary part of the target CSI matrix are respectively input into their respective corresponding target CSI decoders for decoding. In other words, the UE deploys two target CSI decoders at the same time, and can input the compressed coding value of the real part of the target CSI matrix into one of the target CSI decoders, and in parallel input the compressed coding value of the imaginary part of the target CSI matrix into the other target CSI decoder. In the embodiment of the present application, in the parallel decoding method, since two target CSI decoders are deployed at the same time, the real and imaginary parts of the target CSI matrix can be decoded synchronously, making the decoding speed of the parallel decoding method faster.

[0198] It can be understood that if the compressed coded value of the real part corresponds to the first target CSI decoder, then the compressed coded value of the imaginary part corresponds to the second target CSI decoder, or, if the compressed coded value of the real part corresponds to the second target CSI decoder, then the compressed coded value of the imaginary part corresponds to the first target CSI decoder. Optionally, if the first target CSI decoder is trained using the compressed coded value of the real part of the sample CSI matrix, then the compressed coded value of the real part of the target CSI matrix corresponds to the first target CSI decoder, and the compressed coded value of the imaginary part of the target CSI matrix corresponds to the second target CSI decoder. If the first target CSI decoder is trained using the compressed coded value of the imaginary part of the sample CSI matrix, then the compressed coded value of the imaginary part of the target CSI matrix corresponds to the first target CSI decoder, and the compressed coded value of the real part of the target CSI matrix corresponds to the second target CSI decoder.

[0199] In the CSI decoding device provided in the embodiment of the present application, the delay domain convolution step size of the first composite convolution layer in the first target CSI compression encoder deployed on the UE is smaller than the angle domain convolution step size. Since the correlation of the target CSI matrix in the angle domain is strong and the correlation in the delay domain is low, the first composite convolution layer can be used to perform more compression on the angle domain dimension with strong correlation, and less compression on the delay domain dimension with low correlation. Accordingly, the delay domain convolution step size of the Nth composite deconvolution layer in the first target CSI decoder deployed on the network device is smaller than the angle domain convolution step size, so that the Nth composite deconvolution layer can be used to perform more recovery on the angle domain dimension with strong correlation, and less recovery on the delay domain dimension with low correlation, thereby achieving accurate decoding of the compressed coding value of the target CSI matrix.

[0200] Figure 14 This is a schematic diagram of the structure of a CSI decoding device provided in an embodiment of the present application. Figure 14 As shown, the CSI decoding device 400 includes: a receiving module 41, a decoding module 42, a training module 43, an updating module 44 and an inverse DFT module 45.

[0201] The receiving module 41 is configured to receive a compressed coding value corresponding to a target CSI matrix in a delay angle domain sent by a UE;

[0202] A decoding module 42 is used to decode the compressed coding value based on a first target CSI decoder to generate an estimated value of the target CSI matrix, wherein the first target CSI decoder includes N compound deconvolution layers and one fully connected layer, each compound deconvolution layer includes a delay domain convolution step and an angle domain convolution step, the delay domain convolution step of the Nth compound deconvolution layer among the N compound deconvolution layers is smaller than the angle domain convolution step of the Nth compound deconvolution layer, and N is a positive integer.

[0203] Optionally, the receiving module 41 is further used to receive a sample compression coding value sent by the UE, wherein the sample compression coding value is generated by the UE encoding the real part or imaginary part of the sample CSI matrix in the delay angle domain based on the CSI compression encoder.

[0204] Optionally, the training module 43 is configured to train a CSI decoder based on the sample compression coding value to obtain the first target CSI decoder.

[0205] Optionally, the Nth composite deconvolution layer also needs to meet the following conditions:

[0206] The convolution step size of the delay domain of the Nth composite deconvolution layer is smaller than the width of the convolution kernel of the Nth composite deconvolution layer;

[0207] The convolution step size of the angle domain of the Nth composite deconvolution layer is smaller than the number of input channels of the convolution kernel in the Nth composite deconvolution layer.

[0208] Optionally, the updating module 44 is configured to receive an updated compressed coding value of the updated CSI matrix in the delay angle domain sent by the UE, and update the first target CSI decoder based on the updated compressed coding value.

[0209] Optionally, the compressed coded value includes a first compressed coded value of a real part and a second compressed coded value of an imaginary part of the target CSI matrix, wherein the decoding module 42 is further configured to:

[0210] constructing a second target CSI decoder based on the model parameters of the first target CSI decoder;

[0211] The first compressed coded value and the second compressed coded value are respectively input into their corresponding target CSI decoders for decoding.

[0212] Optionally, the decoding module 42 is further configured to:

[0213] If the first target CSI decoder is a decoder trained based on the real part of the sample CSI matrix, inputting the first compressed coded value into the first target CSI decoder for decoding, and inputting the second compressed coded value into the second target CSI decoder for decoding; or,

[0214] If the first target CSI decoder is a decoder trained based on the imaginary part of the sample CSI matrix, the first compressed coded value is input into the second target CSI decoder for decoding, and the second compressed coded value is input into the first target CSI decoder for decoding.

[0215] Optionally, the compressed coded value includes a first compressed coded value of the real part and a second compressed coded value of the imaginary part of the target CSI matrix, wherein the decoding module 42 is further used to input the first compressed coded value and the second compressed coded value into the first target CSI decoder in sequence for decoding.

[0216] Optionally, the process of the first target CSI decoder performing decoding includes:

[0217] Inputting the compressed coded value into a fully connected layer to output a vector, and reconstructing the vector to generate a first tensor;

[0218] Output the first tensor to the first composite deconvolution layer to obtain a second tensor;

[0219] The tensor output by the previous compound deconvolution layer is used as the input of the next compound deconvolution layer, until the Nth compound deconvolution layer, wherein the output of the Nth compound deconvolution layer is the estimated value of the target CSI.

[0220] Optionally, each of the N composite deconvolution layers includes a deconvolution layer, a batch normalization layer, and an activation function layer.

[0221] Optionally, the decoding module 42 is further configured to perform the same zero-padding process on the input of each composite deconvolution layer.

[0222] Optionally, the inverse DFT module 45 is configured to perform an inverse DFT on the estimated value of the target CSI matrix to obtain a space-frequency domain CSI matrix.

[0223] In the CSI decoding device provided in the embodiment of the present application, the delay domain convolution step size of the Nth composite deconvolution layer in the first target CSI decoder deployed on the network device is smaller than the angle domain convolution step size, so that the Nth composite deconvolution layer can perform more recovery on the angle domain dimension with strong correlation and less recovery on the delay domain dimension with low correlation, thereby achieving accurate decoding of the compressed coding value of the target CSI matrix.

[0224] According to an embodiment of the present application, the present application also provides a communication device and a readable storage medium.

[0225] like Figure 15 As shown, the communication device includes: one or more processors 1610, a memory 1620, and interfaces for connecting various components, including high-speed interfaces and low-speed interfaces. The various components are connected to each other using different buses and can be installed on a common mainboard or installed in other ways as needed. The processor can process instructions executed in the communication device, including instructions stored in or on the memory to display graphical information of a GUI on an external input / output device (such as a display device coupled to the interface). In other embodiments, if necessary, multiple processors and / or multiple buses can be used together with multiple memories and multiple memories. Similarly, multiple communication devices can be connected, and each device provides some necessary operations (for example, as a server array, a group of blade servers, or a multi-processor system). Figure 15 A processor 1610 is taken as an example.

[0226] Memory 1620 is the non-transitory computer-readable storage medium provided in this application. The memory stores instructions executable by at least one processor, causing the at least one processor to perform the CSI encoding method or CSI decoding method provided in this application. The non-transitory computer-readable storage medium of this application stores computer instructions for causing a computer to perform the CSI encoding method or CSI decoding method provided in this application.

[0227] Memory 1620, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs, non-transitory computer-executable programs, and modules, such as the program instructions / modules corresponding to the method for using reference time information in the embodiments of the present application. Processor 1610 executes the non-transitory software programs, instructions, and modules stored in memory 1620 to execute various server functional applications and data processing, namely, to implement the CSI encoding method or CSI decoding method in the above-mentioned method embodiments.

[0228] Memory 1620 may include a program storage area and a data storage area. The program storage area may store an operating system and applications required for at least one function; the data storage area may store data generated based on the use of the positioning communication device. Furthermore, memory 1620 may include high-speed random access memory and non-volatile memory, such as at least one disk storage device, flash memory device, or other non-volatile solid-state storage device. Optionally, memory 1620 may include memory remotely located relative to processor 1610, and such remote memory may be connected to the positioning communication device via a network. Examples of such networks include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0229] The communication device may further include: an input device 1630 and an output device 1640. The processor 1610, the memory 1620, the input device 1630 and the output device 1640 may be connected via a bus or other means. Figure 15 The bus connection is taken as an example.

[0230] Input device 1630 can receive input digital or character information and generate key signal input related to user settings and function control of the positioning communication device, such as input devices such as a touch screen, a keypad, a mouse, a trackpad, a touch pad, a pointer, one or more mouse buttons, a trackball, a joystick, etc. Output device 1640 can include a display device, an auxiliary lighting device (e.g., an LED), and a tactile feedback device (e.g., a vibration motor). The display device can include, but is not limited to, a liquid crystal display (LCD), a light emitting diode (LED) display, and a plasma display. In some embodiments, the display device can be a touch screen.

[0231] Various implementations of the systems and techniques described herein can be realized in digital electronic circuit systems, integrated circuit systems, dedicated ASICs (application specific integrated circuits), computer hardware, firmware, software, and / or combinations thereof. These various implementations can include being implemented in one or more computer programs that can be executed and / or interpreted on a programmable system that includes at least one programmable processor, which can be a special purpose or general purpose programmable processor that can receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.

[0232] These computer programs (also referred to as programs, software, software applications, or code) include machine instructions for a programmable processor and can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. As used herein, the terms "machine-readable medium" and "computer-readable medium" refer to any computer program product, apparatus, and / or device (e.g., a magnetic disk, an optical disk, a memory, a programmable logic device (PLD)) for providing machine instructions and / or data to a programmable processor, including a machine-readable medium that receives machine instructions as a machine-readable signal. The term "machine-readable signal" refers to any signal for providing machine instructions and / or data to a programmable processor.

[0233] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user can provide input to the computer. Other types of devices can also be used to provide interaction with the user; for example, the feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including acoustic input, voice input, or tactile input).

[0234] The systems and techniques described herein can be implemented in a computing system that includes back-end components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes front-end components (e.g., a user computer having a graphical user interface or a web browser through which a user can interact with implementations of the systems and techniques described herein), or a computing system that includes any combination of such back-end components, middleware components, or front-end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include a local area network (LAN), a wide area network (WAN), and the Internet.

[0235] Computer systems may include clients and servers. A client and server are generally remote from each other and typically interact through a communication network. The client and server relationship arises through computer programs running on the respective computers and having a client-server relationship to each other.

[0236] Those skilled in the art will understand that all or part of the steps in the method of the above embodiment can be completed by instructing related hardware through a program, and the program can be stored in a computer-readable storage medium. When the program is executed, it includes one or a combination of the steps of the method embodiment.

[0237] In addition, the functional units in the various embodiments of the present invention may be integrated into a single processing module, or each unit may exist physically separately, or two or more units may be integrated into a single module. The aforementioned integrated modules may be implemented in the form of hardware or in the form of software functional modules. If the integrated modules are implemented in the form of software functional modules and sold or used as independent products, they may also be stored in a computer-readable storage medium.

[0238] The storage medium mentioned above can be a read-only memory, a magnetic disk or an optical disk, etc.

[0239] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A CSI encoding method, characterized in that: Applicable to user equipment UE, the method includes: The target CSI matrix in the delay angle domain is encoded based on the first target CSI compression encoder to generate a compressed coding value, wherein the first target CSI compression encoder includes N compound convolutional layers and one fully connected layer, each compound convolutional layer includes a delay domain convolution step and an angle domain convolution step, the delay domain convolution step of the first compound convolution layer among the N compound convolution layers is smaller than the angle domain convolution step of the first compound convolution layer, and N is a positive integer.

2. The CSI encoding method according to claim 1, wherein: Also includes: Based on the real part or the imaginary part of the sample CSI matrix in the delay angle domain, the CSI compression encoder is trained to obtain the first target CSI compression encoder.

3. The CSI encoding method according to claim 1 or 2, characterized in that: The first composite convolutional layer also needs to meet the following conditions: The delay domain convolution step size of the first composite convolutional layer is smaller than the width of the convolution kernel in the first composite convolutional layer; The angle domain convolution step size of the first composite convolution layer is smaller than the number of input channels of the convolution kernel in the first composite convolution layer.

4. The CSI encoding method according to claim 1 or 2, characterized in that: Also includes: An updated CSI matrix in the delay angle domain is obtained, and the first target CSI compression encoder is updated based on the updated CSI matrix.

5. The CSI encoding method according to claim 2, wherein: The encoding of the target CSI matrix in the delay angle domain based on the first target CSI compression encoder includes: constructing a second target CSI compression encoder based on the model parameters of the first target CSI compression encoder; Extracting the real and imaginary parts of the target CSI matrix; The real part and the imaginary part of the target CSI matrix are respectively input into corresponding target CSI compression encoders for encoding.

6. The CSI encoding method according to claim 5, characterized in that Inputting the real part and the imaginary part of the target CSI matrix into respective corresponding target CSI compression encoders for encoding includes: If the first target CSI compression encoder is an encoder trained based on the real part of the sample CSI matrix, the real part of the target CSI matrix is input into the first target CSI compression encoder for encoding, and the imaginary part of the target CSI matrix is input into the second target CSI compression encoder for encoding; or If the first target CSI compression encoder is an encoder trained based on the imaginary part of the sample CSI matrix, the imaginary part of the target CSI matrix is input into the first target CSI compression encoder for encoding, and the real part of the target CSI matrix is input into the second target CSI compression encoder for encoding.

7. The CSI encoding method according to claim 1 or 2, characterized in that: The encoding of the target CSI matrix in the delay angle domain based on the first target CSI compression encoder includes: The real part and the imaginary part of the target CSI matrix are extracted, and the real part and the imaginary part are sequentially input into the first target CSI compression encoder for encoding.

8. The CSI encoding method according to claim 1, wherein: The encoding process of the first target CSI compression encoder includes: Inputting the real part and / or imaginary part of the target CSI matrix into the first composite convolutional layer to obtain a tensor output by the first composite convolutional layer; Using the tensor output by the previous composite convolutional layer as the input of the next composite convolutional layer, until the Nth composite convolutional layer; The tensor output by the Nth composite convolutional layer is reconstructed, and the reconstructed vector is input into a fully connected layer to generate the compressed coded value.

9. The CSI encoding method according to claim 1 or 8, characterized in that: Each of the N composite convolutional layers includes a convolutional layer, a batch normalization layer, and an activation function layer.

10. The CSI encoding method according to claim 1 or 8, characterized in that: Also includes: The input of each composite convolutional layer is padded with zeros.

11. The CSI encoding method according to claim 1, wherein: Also includes: Obtain a CSI matrix in the space-frequency domain, perform a two-dimensional discrete Fourier transform (DFT) on the CSI matrix to obtain the target CSI matrix, and retain non-zero row principal values.

12. The CSI encoding method according to claim 1, wherein: Also includes: The compressed encoded value is sent to a network device.

13. A CSI decoding method, characterized in that: Applicable to a network device, the method includes: receiving a compressed coding value corresponding to a target CSI matrix in a delay angle domain sent by a user equipment UE; The compressed coding value is decoded based on a first target CSI decoder to generate an estimated value of the target CSI matrix, wherein the first target CSI decoder includes N compound deconvolution layers and one fully connected layer, each compound deconvolution layer includes a delay domain convolution step and an angle domain convolution step, the delay domain convolution step of the Nth compound deconvolution layer among the N compound deconvolution layers is smaller than the angle domain convolution step of the Nth compound deconvolution layer, and N is a positive integer.

14. The CSI decoding method according to claim 13, wherein: Also includes: receiving a sample compression coded value sent by the UE, wherein the sample compression coded value is generated by the UE encoding a real part or an imaginary part of a sample CSI matrix in a delay angle domain based on a CSI compression encoder; A CSI decoder is trained based on the sample compression coding value to obtain the first target CSI decoder.

15. The CSI decoding method according to claim 13 or 14, characterized in that: The Nth composite deconvolution layer also needs to meet the following conditions: The convolution step size of the delay domain of the Nth composite deconvolution layer is smaller than the width of the convolution kernel of the Nth composite deconvolution layer; The convolution step size of the angle domain of the Nth composite deconvolution layer is smaller than the number of input channels of the convolution kernel in the Nth composite deconvolution layer.

16. The CSI decoding method according to claim 13 or 14, characterized in that: Also includes: Receive an updated compressed coding value of the updated CSI matrix in the delay angle domain sent by the UE, and update the first target CSI decoder based on the updated compressed coding value.

17. The CSI decoding method according to claim 14, wherein: The compressed coded value includes a first compressed coded value of a real part and a second compressed coded value of an imaginary part of the target CSI matrix, wherein decoding the compressed coded value based on the first target CSI decoder includes: constructing a second target CSI decoder based on the model parameters of the first target CSI decoder; The first compressed coded value and the second compressed coded value are respectively input into their corresponding target CSI decoders for decoding.

18. The CSI decoding method according to claim 17, wherein: Inputting the first compressed coded value and the second compressed coded value into respective corresponding target CSI decoders for decoding includes: If the first target CSI decoder is a decoder trained based on the real part of the sample CSI matrix, inputting the first compressed coded value into the first target CSI decoder for decoding, and inputting the second compressed coded value into the second target CSI decoder for decoding; or, If the first target CSI decoder is a decoder trained based on the imaginary part of the sample CSI matrix, the first compressed coded value is input into the second target CSI decoder for decoding, and the second compressed coded value is input into the first target CSI decoder for decoding.

19. The CSI decoding method according to claim 13 or 14, characterized in that: The compressed coded value includes a first compressed coded value of a real part and a second compressed coded value of an imaginary part of the target CSI matrix, wherein decoding the compressed coded value based on the first target CSI decoder includes: The first compressed coded value and the second compressed coded value are sequentially input into the first target CSI decoder for decoding.

20. The CSI decoding method according to claim 13, wherein: The process of decoding by the first target CSI decoder includes: Inputting the compressed coded value into a fully connected layer to output a vector, and reconstructing the vector to generate a first tensor; Output the first tensor to the first composite deconvolution layer to obtain a second tensor; The tensor output by the previous compound deconvolution layer is used as the input of the next compound deconvolution layer, until the Nth compound deconvolution layer, wherein the output of the Nth compound deconvolution layer is the estimated value of the target CSI.

21. The CSI decoding method according to claim 13 or 20, characterized in that: Each of the N composite deconvolution layers includes a deconvolution layer, a batch normalization layer, and an activation function layer.

22. The CSI decoding method according to claim 13 or 20, characterized in that: Also includes: The input of each composite deconvolution layer is padded with zeros.

23. The CSI decoding method according to claim 13, wherein: Also includes: Perform an inverse DFT on the estimated value of the target CSI matrix to obtain a space-frequency domain CSI matrix.

24. A CSI encoding device, characterized in that: The device comprises: An encoding module is used to encode the target CSI matrix in the delay angle domain based on a first target CSI compression encoder to generate a compressed coding value, wherein the first target CSI compression encoder includes N compound convolutional layers and a fully connected layer, each compound convolutional layer includes a delay domain convolution step and an angle domain convolution step, the delay domain convolution step of the first compound convolution layer among the N compound convolution layers is smaller than the angle domain convolution step of the first compound convolution layer, and N is a positive integer.

25. The CSI encoding device according to claim 24, characterized in that Also includes: A training module is used to train a CSI compression encoder based on the real part or imaginary part of the sample CSI matrix in the delay angle domain to obtain the first target CSI compression encoder.

26. The CSI encoding device according to claim 24 or 25, characterized in that The first composite convolutional layer also needs to meet the following conditions: The delay domain convolution step size of the first composite convolutional layer is smaller than the width of the convolution kernel in the first composite convolutional layer; The angle domain convolution step size of the first composite convolution layer is smaller than the number of input channels of the convolution kernel in the first composite convolution layer.

27. The CSI encoding device according to claim 24 or 25, characterized in that Also includes: An updating module is configured to obtain an updated CSI matrix in the delay angle domain, and update the first target CSI compression encoder based on the updated CSI matrix.

28. The CSI encoding device according to claim 25, characterized in that The encoding module is further used for: constructing a second target CSI compression encoder based on the model parameters of the first target CSI compression encoder; Extracting the real and imaginary parts of the target CSI matrix; The real part and the imaginary part of the target CSI matrix are respectively input into corresponding target CSI compression encoders for encoding.

29. The CSI encoding device according to claim 28, wherein: The encoding module is further used for: If the first target CSI compression encoder is an encoder trained based on the real part of the sample CSI matrix, inputting the real part of the target CSI matrix into the first target CSI compression encoder for encoding, and inputting the imaginary part of the target CSI matrix into the second target CSI compression encoder for encoding; or, If the first target CSI compression encoder is an encoder trained based on the imaginary part of the sample CSI matrix, the imaginary part of the target CSI matrix is input into the first target CSI compression encoder for encoding, and the real part of the target CSI matrix is input into the second target CSI compression encoder for encoding.

30. The CSI encoding device according to claim 24 or 25, characterized in that The encoding module is further used for: The real part and the imaginary part of the target CSI matrix are extracted, and the real part and the imaginary part are sequentially input into the first target CSI compression encoder for encoding.

31. The CSI encoding device according to claim 24, wherein: The encoding process of the first target CSI compression encoder includes: Inputting the real part and / or imaginary part of the target CSI matrix into the first composite convolutional layer to obtain a tensor output by the first composite convolutional layer; Using the tensor output by the previous composite convolutional layer as the input of the next composite convolutional layer, until the Nth composite convolutional layer; The tensor output by the Nth composite convolutional layer is reconstructed, and the reconstructed vector is input into a fully connected layer to generate the compressed coded value.

32. The CSI encoding device according to claim 24 or 31, characterized in that Each of the N composite convolutional layers includes a convolutional layer, a batch normalization layer, and an activation function layer.

33. The CSI encoding device according to claim 24 or 31, characterized in that The encoding module is further configured to perform the same zero-padding processing on the input of each composite convolutional layer.

34. The CSI encoding device according to claim 24, characterized in that Also includes: The conversion module is used to obtain a CSI matrix in the space-frequency domain, perform a two-dimensional discrete Fourier transform (DFT) on the CSI matrix to obtain the target CSI matrix, and retain the main values of non-zero rows.

35. The CSI encoding device according to claim 24, characterized in that Also includes: The sending module is used to send the compressed code value to the network device.

36. A CSI decoding device, characterized in that: The device comprises: A receiving module, configured to receive a compressed coded value corresponding to a target CSI matrix in a delay angle domain sent by a UE; A decoding module is configured to decode the compressed coding value based on a first target CSI decoder to generate an estimated value of the target CSI matrix, wherein the first target CSI decoder includes N compound deconvolution layers and one fully connected layer, each compound deconvolution layer includes a delay domain convolution step and an angle domain convolution step, the delay domain convolution step of an Nth compound deconvolution layer among the N compound deconvolution layers is smaller than the angle domain convolution step of the Nth compound deconvolution layer, and N is a positive integer.

37. The CSI decoding apparatus according to claim 36, wherein: Also includes: The receiving module is further configured to receive a sample compression code value sent by the UE, wherein the sample compression code value is generated by the UE encoding a real part or an imaginary part of a sample CSI matrix in the delay angle domain based on a CSI compression encoder; A training module is configured to train a CSI decoder based on the sample compression coding value to obtain the first target CSI decoder.

38. The CSI decoding apparatus according to claim 36 or 37, wherein: The Nth composite deconvolution layer also needs to meet the following conditions: The convolution step size of the delay domain of the Nth composite deconvolution layer is smaller than the width of the convolution kernel of the Nth composite deconvolution layer; The convolution step size of the angle domain of the Nth composite deconvolution layer is smaller than the number of input channels of the convolution kernel in the Nth composite deconvolution layer.

39. The CSI decoding apparatus according to claim 36 or 37, wherein: Also includes: An updating module is configured to receive an updated compressed coding value of an updated CSI matrix in a delay angle domain sent by a UE, and update the first target CSI decoder based on the updated compressed coding value.

40. The CSI decoding apparatus according to claim 37, wherein: The compressed coded value includes a first compressed coded value of a real part and a second compressed coded value of an imaginary part of the target CSI matrix, wherein the decoding module is further configured to: constructing a second target CSI decoder based on the model parameters of the first target CSI decoder; The first compressed coded value and the second compressed coded value are respectively input into their corresponding target CSI decoders for decoding.

41. The CSI decoding apparatus according to claim 40, wherein: The decoding module is further used for: If the first target CSI decoder is a decoder trained based on the real part of the sample CSI matrix, inputting the first compressed coded value into the first target CSI decoder for decoding, and inputting the second compressed coded value into the second target CSI decoder for decoding; or, If the first target CSI decoder is a decoder trained based on the imaginary part of the sample CSI matrix, the first compressed coded value is input into the second target CSI decoder for decoding, and the second compressed coded value is input into the first target CSI decoder for decoding.

42. The CSI decoding apparatus according to claim 36 or 37, wherein: The compressed coded value includes a first compressed coded value of a real part and a second compressed coded value of an imaginary part of the target CSI matrix, wherein the decoding module is further configured to: The first compressed coded value and the second compressed coded value are sequentially input into the first target CSI decoder for decoding.

43. The CSI decoding apparatus according to claim 36, wherein: The process of decoding by the first target CSI decoder includes: Inputting the compressed coded value into a fully connected layer to output a vector, and reconstructing the vector to generate a first tensor; Output the first tensor to the first composite deconvolution layer to obtain a second tensor; The tensor output by the previous compound deconvolution layer is used as the input of the next compound deconvolution layer, until the Nth compound deconvolution layer, wherein the output of the Nth compound deconvolution layer is the estimated value of the target CSI.

44. The CSI decoding apparatus according to claim 36 or 43, wherein: Each of the N composite deconvolution layers includes a deconvolution layer, a batch normalization layer, and an activation function layer.

45. The CSI decoding apparatus according to claim 36 or 43, wherein: The decoding module is further configured to perform the same zero-padding process on the input of each composite deconvolution layer.

46. The CSI decoding apparatus according to claim 36, wherein: Also includes: The inverse DFT module is configured to perform an inverse DFT on the estimated value of the target CSI matrix to obtain a space-frequency domain CSI matrix.

47. A communication device, characterized in that include: at least one processor; as well as a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method according to any one of claims 1 to 23.

48. A computer storage medium, characterized in that The computer storage medium stores computer-executable instructions, and after the computer-executable instructions are executed by the processor, the method according to any one of claims 1 to 23 can be implemented.

Images