CSI feedback method and device based on important value transmission and non-important value generation

Abstract

The invention discloses a CSI (Channel State Information) feedback method and device based on important value transmission and non-important value generation, and relates to the technical field of channel state feedback, and the method comprises the following steps: a UE (User Equipment) side obtains a CSI matrix, selects an algorithm based on important elements, scores the importance of elements in the CSI matrix, sorts the elements in a descending order based on the importance score values, and feeds back the CSI matrix to the UE side; selecting Top-k important elements and an index matrix of the important elements according to a sorting result; the important elements are coded by adopting a self-adaptive quantization coding code length algorithm, an index matrix of the important elements is coded by using an improved Huffman codebook, and coding results are combined and then transmitted to a BS side; and the BS side dequantizes the first coding result to obtain a first decoding result, decodes the second coding result to obtain a second decoding result, combines the first decoding result and the second decoding result to obtain a reconstructed CSI matrix, inputs the reconstructed CSI matrix into a pre-established generator model, and outputs to obtain a CSI matrix reconstruction result.

Description

Technical Field

[0001] The present invention relates to the technical field of channel state feedback, and specifically to a CSI feedback method and device based on the transmission of important values and the generation of unimportant values. Background Art

[0002] In the current large-scale MIMO system with frequency-division duplexing (FDD), the CSI (channel state information) generated on the UE (user equipment) side needs to be compressed and then transmitted to the BS (base station) side over the wireless link. It is required to obtain a high CSI reconstruction accuracy at the BS side with as little communication overhead as possible.

[0003] Existing solutions mainly use deep learning models to achieve the compression and reconstruction of the CSI matrix. The current main deep learning model is the model based on the autoencoder (AE), that is, an encoder is deployed on the UE side to perform matrix dimensionality reduction on the CSI matrix to obtain an intermediate codeword, and the intermediate codeword is quantized and encoded, and after being turned into a binary sequence, it is transmitted to the BS side through the wireless channel. On the BS side, after the received binary sequence is decoded and dequantized, a decoder is deployed, and the dimension of the matrix is expanded through the decoder to restore the dimension of the CSI matrix before compression, and this is used as the result of CSI reconstruction. However, the trained autoencoder model has obvious generalization problems. That is, when the statistical distribution of the CSI matrix samples processed by the system is quite different from the statistical distribution of the data used to train the autoencoder model, the feedback accuracy of the system will decrease significantly. Summary of the Invention

[0004] To solve the deficiencies mentioned in the above background art, the purpose of the present invention is to provide a CSI feedback method and device based on the transmission of important values and the generation of unimportant values.

[0005] In a first aspect, the purpose of the present invention can be achieved through the following technical solutions: A CSI feedback method based on the transmission of important values and the generation of unimportant values, the method comprising the following steps:

[0006] The UE side obtains a CSI matrix, scores the importance of the elements in the CSI matrix based on an important element selection algorithm, sorts the elements in descending order according to the importance score values, and selects the top-k important elements and the index matrix of the important elements according to the sorting result;

[0007] The important elements are encoded using an adaptive quantization coding length algorithm to obtain a first coding result, the index matrix of the important elements is encoded using an improved Huffman codebook to obtain a second coding result, and the first coding result and the second coding result are combined and then transmitted to the BS side;

[0008] The BS side dequantizes the first coding result to obtain the first decoding result, decodes the second coding result to obtain the second decoding result, combines the first decoding result and the second decoding result to obtain a reconstructed CSI matrix, and inputs the reconstructed CSI matrix into a pre-established generator model to output a CSI matrix reconstruction result.

[0009] Combined with the first aspect, in some implementation manners of the first aspect, the method further includes: the calculation process of performing importance scoring on the elements in the CSI matrix based on the important element selection algorithm is as follows:

[0010] The importance score consists of two dimensions, namely the amplitude size and the difference size from adjacent elements;

[0011] Among them, the importance score score i,j , where i and j respectively represent the row index and column index of the element in the CSI matrix, and the calculation formula is as follows:

[0012] score i,j =a i,j +d(i,j),#(1)

[0013] In the formula, a i,j is the amplitude size, and d(i,j) is the difference size from adjacent elements.

[0014] Combined with the first aspect, in some implementation manners of the first aspect, the method further includes: the calculation formulas for the amplitude size a i,j and the difference size d(i,j) from adjacent elements are as follows:

[0015]

[0016] Among them, h i,j is the element in the i-th row and j-th column of the CSI matrix H, real(·) represents the real part of a complex value, and imag(·) represents the imaginary part of a complex value.

[0017] Combined with the first aspect, in some implementation manners of the first aspect, the method further includes: the process of performing descending order sorting based on the importance scoring values:

[0018] Select the k elements with the highest scores, record the index information, represent these k complex elements by real part and imaginary part, and then arrange them in the order of index sequence to form an importance value vector v im . For the index information of the importance value vector, construct an index matrix M.

[0019] In combination with the first aspect, in certain implementations of the first aspect, the method further includes: the process of encoding the index matrix of important elements using an improved Huffman codebook:

[0020] For the compression encoding of the index matrix M, an improved Huffman encoding codebook C is introduced to encode the relative positions of each 1 in M. First, the relative positions of 1s in M are counted to form a relative position vector v p , and then C is used to encode each element in v p to form binary codewords, and then the binary codewords of all elements are concatenated to form a binary string v for encoding M i ;

[0021] The steps include: relative position acquisition, encoding the relative positions using an improved Huffman codebook, and recovering the index matrix from the encoded binary string.

[0022] In combination with the first aspect, in certain implementations of the first aspect, the method further includes: the process of encoding important elements using an adaptive quantization encoding code length algorithm:

[0023] For the quantization, encoding, and decoding of the important value vector v im , first, the important values are quantized, and then encoded. When encoding, two lengths of code lengths q L and q S are used for encoding, where the shorter code length q S is dynamically adjusted according to the important value vector to be transmitted;

[0024] The steps include: quantizing the important values, determining the short code length using an adaptive method, encoding the important values, and decoding the important values.

[0025] In combination with the first aspect, in certain implementations of the first aspect, the method further includes: the process of inputting the reconstructed CSI matrix into a pre-established generator model and outputting the CSI matrix reconstruction result:

[0026] Initialize a fully connected layer, input H m , multiply the obtained output element-wise with M to obtain a masked feature matrix F m , where the feature vectors at the positions of non-important elements are all-zero vectors;

[0027] Initialize a padding feature vector v p , copy it, and fill it into all positions of non-important elements. All the feature vectors at the positions of non-important elements in F m are replaced with v p to obtain a padded feature matrix F p ;

[0028] Process the sequence of feature vectors using multiple Transformer layers. For the two-dimensional matrix F composed of feature vectors p , first cut it column by column. Each column of feature vectors is regarded as a sequence of feature vectors. All the sequences of feature vectors are regarded as a batch and input into 8 cascaded Transformer layers for processing to obtain the processed feature vector matrix F pc , and process F p in the same way. Split it row by row, and then process it to obtain the feature vector matrix F pl , and splice F pc and F pl along the dimension of the feature vectors to obtain the feature matrix F processed by the Transformer layer out ;

[0029] Initialize a fully connected layer. Input F out , and output a matrix H with the same dimension as the CSI matrix out . Replace the positions of the important elements in H out with the values at the corresponding positions in H m to obtain the final CSI reconstruction result H

[0030] In a second aspect, to achieve the above object, the present invention discloses a CSI feedback device based on important value transmission and non-important value generation, including:

[0031] A matrix splitting module, configured to obtain a CSI matrix on the UE side, score the importance of the elements in the CSI matrix based on an important element selection algorithm, sort them in descending order based on the importance score values, and select the top-k important elements and the index matrix of the important elements according to the sorting result;

[0032] An element encoding module, configured to encode the important elements using an adaptive quantization coding length algorithm to obtain a first coding result, encode the index matrix of the important elements using an improved Huffman codebook to obtain a second coding result, and merge the first coding result and the second coding result and transmit them to the BS side;

[0033] A matrix reconstruction module, configured to dequantize the first coding result on the BS side to obtain a first decoding result, decode the second coding result to obtain a second decoding result, merge the first decoding result and the second decoding result to obtain a reconstructed CSI matrix, and input the reconstructed CSI matrix into a pre-established generator model to output a CSI matrix reconstruction result

[0034] In another aspect of the present invention, in order to achieve the above object, a terminal device is disclosed, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor. The memory stores a computer program capable of running on the processor. When the processor loads and executes the computer program, a CSI feedback method based on important value transmission and non-important value generation as described above is adopted.

[0035] In yet another aspect of the present invention, in order to achieve the above object, a computer-readable storage medium is disclosed. The computer-readable storage medium stores a computer program. When the computer program is loaded and executed by a processor, a CSI feedback method based on important value transmission and non-important value generation as described above is adopted.

[0036] Advantages of the present invention:

[0037] Compared with the CSI feedback system in which deep learning models are deployed on both the terminal side and the base station side, the system for important element transmission and non-important element generation only deploys a deep learning model on the base station side and uses a general algorithm for all scenarios on the terminal side. This improves the generalization of the entire CSI feedback model and can ensure performance in multiple channel scenarios. At the same time, compared with other CSI feedback systems based on partial value transmission, this system uses a Transformer model on the base station side to predict those non-important elements that are not transmitted based on the transmitted important elements, improving the reconstruction accuracy of CSI feedback. Description of the Drawings

[0038] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, for those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings;

[0039] Figure 1 It is a schematic flowchart of the method of the present invention;

[0040] Figure 2 It is a schematic diagram of the method for obtaining the values and indexes of important elements of the present invention;

[0041] Figure 3 It is a schematic flowchart of the encoding process of the index matrix M of the present invention;

[0042] Figure 4 It is a schematic diagram of the process of short code length selection of the present invention;

[0043] Figure 5 It is a schematic flowchart of the implementation process of the important element quantization encoding and decoding method of the present invention;

[0044] Figure 6 It is a schematic diagram of the present invention using a generative model to predict unimportant elements;

[0045] Figure 7 It is a comparison diagram of the present invention with an autoencoder-based CSI feedback scheme in terms of communication overhead and feedback performance;

[0046] Figure 8 It is a schematic diagram of the system structure of the present invention. Detailed implementation manners

[0047] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0048] Embodiment 1:

[0049] Next, the related terms involved in the embodiments of the present application will be introduced:

[0050] CSI: Channel State Information (abbreviated as CSI), a term in the field of wireless communication, which is the channel attribute of a communication link. It describes the attenuation factor of the signal on each transmission path, that is, the value of each element in the channel gain matrix H, such as signal scattering, environmental attenuation (fading, multipath fading or shadowing fading), distance attenuation (power decay of distance), etc. CSI can enable the communication system to adapt to the current channel conditions and provide guarantee for high-reliability and high-rate communication in a multi-antenna system.

[0051] As Figure 1 shown, a CSI feedback method based on important value transmission and unimportant value generation, the method includes the following steps:

[0052] The UE side obtains a CSI matrix, based on an important element selection algorithm, scores the importance of the elements in the CSI matrix, sorts them in descending order based on the importance score values, and selects the Top-k important elements and the index matrix of the important elements according to the sorting results;

[0053] The calculation process of scoring the importance of the elements in the CSI matrix based on the important element selection algorithm is as follows:

[0054] The first step in selecting important elements is to calculate the importance score score for each element in the CSI matrix i,j , where i and j represent the row index and column index of the element in the CSI matrix respectively. Calculating the importance score consists of two dimensions, namely the magnitude size and the difference size from adjacent elements. The reason for considering the first dimension is that in the channel matrix, elements with large magnitudes often represent scattering paths with strong energy in that resolution interval. Therefore, the larger the magnitude, the higher the score for this dimension. The reason for considering the second dimension is that the greater the difference between an element and its adjacent elements, the more irregular its appearance, and the greater the difficulty in predicting it using the generator on the BS side later. So, for elements with a large difference from adjacent elements, a higher importance score should be given so that they can be transmitted and the prediction difficulty of the generator on the BS side can be reduced. To sum up, the calculation formula for the importance score is:

[0055] score i,j = a i,j + d(i,j), #(1)

[0056] In the formula, a i,j is the magnitude size, and d(i,j) is the difference size from adjacent elements.

[0057] The calculation formulas for the magnitude size a i,j and the difference size d(i,j) from adjacent elements are as follows:

[0058]

[0059] where h i,j is the element in the i-th row and j-th column of the CSI matrix H, real(·) represents the real part of a complex value, and imag(·) represents the imaginary part of a complex value.

[0060] The process of sorting in descending order based on the importance score values:

[0061] As Figure 2 shown, select the top k elements with the highest scores, record the index information, represent these k complex elements in terms of real and imaginary parts, and then arrange them in the order of the indices to form the importance value vector v im . For the index information of the importance value vector, construct the index matrix M.

[0062] For important elements, use the adaptive quantization coding length algorithm for coding to obtain the first coding result, use the improved Huffman codebook for coding the index matrix of important elements to obtain the second coding result, and merge the first coding result and the second coding result and then transmit them to the BS side;

[0063] The process of encoding the index matrix of important elements using an improved Huffman codebook:

[0064] For the compression encoding of the index matrix M, an improved Huffman encoding codebook C is introduced to encode the relative positions of each 1 in M. First, the relative positions of 1s in M are counted to form a relative position vector v p , and then C is used to encode p each element in v i to form binary codewords, and then the binary codewords of all elements are concatenated to form a binary string v for encoding M

[0065] The steps include: obtaining relative positions

[0066] The two-dimensional index matrix M is expanded into a one-dimensional vector in a "snake-like" order starting from the upper left corner (the first row and the first column). The so-called "snake-like" means that the next element after the last element in the first row is the last element in the second row, and then the second row is traversed in reverse order from large to small indexes, and then the next element after the first and element in the second row is the first element in the third row, and then the third row is traversed in ascending order of indexes, and so on. After obtaining the one-dimensional vector obtained by expanding M, the index positions where each "1" element is located are counted, and these positions are concatenated into a relative position vector v p , v p The value of each element in

[0067] Encoding the relative positions using an improved Huffman codebook

[0068] Each element in the relative position vector v p is encoded into a binary codeword using a designed Huffman codebook. This codebook uses variable-length encoding. The smaller the element value, the fewer bits it is actually encoded with. The way to convert the decimal element values in v p into binary codewords is direct conversion, and the variable-length codewords ensure that the first bit of each binary codeword is 1. For example, the binary codeword encoded for the decimal value 8 is 100, and the binary codeword encoded for the decimal value 17 is 1001, and so on. For the convenience of unique decoding, a prefix code needs to be added before each binary codeword. In this way, v pThe encoded string of each element in it is composed of two parts: the prefix code and its own codeword. The above encoding method can be represented by an improved Huffman codebook, as shown in Table 1. Using this encoding method can save a large number of transmission bits compared with directly transmitting the index matrix M. Because M is sparse, the number of "1" elements is small, and the probability of the relative position value taking a small value is large. Therefore, using this variable-length encoding can greatly reduce the number of "0" in the encoded binary sequence and reduce the number of bits used for transmission. Concatenate v p the encoded strings of each element in it to obtain the encoded binary string b i of the two-dimensional index matrix M. The flowchart of this step is as Figure 3 shown.

[0069] Table 1 Improved Huffman codebook

[0070]

[0071] Recover the index matrix from the encoded binary string

[0072] When the base station receives the binary string b i after that, since the used Huffman codebook is a uniquely decodable code, b i can be uniquely restored to the relative position vector v p in the case of no transmission error. And v p can uniquely restore the index matrix M. Thus, the encoding and decoding process of the index matrix is completed.

[0073] Relative position acquisition, encoding the relative position using an improved Huffman codebook, and recovering the index matrix from the encoded binary string.

[0074] The process of encoding important elements using an adaptive quantization encoding code length algorithm:[[]]

[0075] For the quantization, encoding, and decoding of the important value vector v im , first quantize the important values, and then encode them. When encoding, use two lengths of code lengths q L and q S to encode to save transmission bits. Among them, the shorter code length q S is dynamically adjusted according to the important value vector to be transmitted;

[0076] The specific steps are as follows:

[0077] Quantize the important values

[0078] Quantize all the values in the important value vector v im . The quantization uses μ-law quantization with μ = 16. The number of quantization intervals needs to be specified in advance. The number of quantization intervals is related to the longer code length qL is related, i.e., the number of quantization intervals = 2 qL . Therefore, the operation of this step is: after performing μ-law compression on all values in the important value vector, perform fixed-length quantization coding on the absolute values, and uniformly code them with a code length of q L +1. The extra bit is the sign bit and is at the first position of the codeword.

[0079] Use an adaptive method to determine the shorter code length

[0080] After performing quantization coding on all values in v using the unified coding code length im , it is noted that after coding some smaller values, there are many invalid "0"s after the sign bit and before the valid bit (the first "1"). Now consider introducing a second shorter code length q S , and let some small values be coded using this short code length, and then add an indication bit at the very front of the coding string of each value to indicate whether the coding of this value selects the long code length q L or the short code length q S . q S is selected using an adaptive method. Starting from q S = q L -1, traverse the code lengths from long to short, and then traverse each value in v im . If it can be coded using q S , record the number of bits saved when the current value is changed from coding with q L to coding with q S . After traversing all values, count the total number of bits saved. For each q S , compare all the total number of bits saved, and select the q S with the most saved bits as the final q S . The process of selecting q S is as shown in Figure 4 .

[0081] Coding of important values

[0082] After determining the long code length q L and the short code length q S , for each value in the important value vector v im , determine whether it can be coded using the short code length. If not, use the long code length for coding. After the coding is completed, add a bit at the beginning of the codeword string to indicate whether the coding is done using the long code or the short code. Concatenate the coding strings of each value in v im to obtain the coding string b of the important values v .

[0083] Decoding of important values

[0084] At the base station, after receiving the coding string b of the important valuesv After that, first judge its first digit. If it indicates a long code, use the length of the long code to convert the binary into a decimal value. Then continue to detect the current digit and loop this process to solve all the important values after μ-law non-uniform transformation. Then use the inverse transformation of μ-law to obtain the values of the important elements before the transformation.

[0085] After the base station obtains the index matrix and the important value vector, fill in each "1" in the index matrix from the upper left to the lower right in row-major order according to the order of the elements in the important value vector. The schematic diagrams of the quantization encoding and decoding of all important elements are as Figure 5 shown.

[0086] The BS side dequantizes the first coding result to obtain the first decoding result, decodes the second coding result to obtain the second decoding result, merges the first decoding result and the second decoding result to obtain the reconstructed CSI matrix, and inputs the reconstructed CSI matrix into the pre-established generator model to output the CSI matrix reconstruction result.

[0087] Implementation of predicting non-important elements using the generative model

[0088] After obtaining the CSI matrix containing only important elements through important element decoding at the base station use a generative model based on the Transformer structure to complete the values at those "0" positions, that is, predict non-important elements based on important elements. The specific steps are as follows:

[0089] Initialize a fully connected layer, take the input, multiply the obtained output element-wise with M to obtain the masked feature matrix F m , where the feature vectors at the positions of non-important elements are all zero vectors.

[0090] Initialize the padding feature vector v p , copy it, and fill it into all positions of non-important elements, that is, all the feature vectors at the positions of non-important elements in F m are all replaced by v p to obtain the padded feature matrix F p .

[0091] Use multiple Transformer layers to process the sequence of feature vectors. For the two-dimensional matrix F p formed by feature vectors, first cut it column by column. The feature vectors of each column are regarded as a sequence of feature vectors. Regard all the sequences of feature vectors (all columns of F p ) as a batch and input them into 8 cascaded Transformer layers for processing to obtain the processed feature vector matrix Fpc , split F by rows in the same way, and then process to obtain the feature vector matrix F p 。 pl Concatenate F pc and F pl along the dimension of the feature vectors to obtain the feature matrix F processed by the Transformer layer out 。

[0092] Initialize a fully connected layer, with F as the input out , and output a matrix H with the same dimension as the CSI matrix out 。In this matrix, both the unimportant elements and the important elements are the predicted values of the generative network. Actually, the important elements should use the values received by BS. Therefore, replace the positions of the important elements in H out with the values at the corresponding positions to obtain the final CSI reconstruction result The entire processing flow is as shown in Figure 6 。

[0093] Construct a training set. For the learnable parameters mentioned above, use the Mean Square Error (MSE) as the loss function, calculate the loss between the predicted unimportant elements and the true unimportant elements, and use the backpropagation algorithm and the Adam optimizer to update the learnable parameters, so as to optimize the model.

[0094] Example 2: Second, as shown in Figure 7 , in order to prove the beneficial effects of the present invention, the present invention discloses a set of performance comparison results with other existing CSI feedback schemes.

[0095] First, define the evaluation metrics for different CSI feedback schemes. Generally speaking, the evaluation metric for evaluating the similarity between the reconstructed CSI matrix and the original matrix is the Normalized Mean-Square Error (NMSE). The definition formula of this metric is

[0096]

[0097] where N test represents the number of samples participating in the calculation, H i represents the original CSI matrix, represents the reconstructed CSI matrix.

[0098] Secondly, three currently widely adopted CSI feedback networks based on autoencoders are implemented, namely CRNet, ACRNet, and TransNet. For these three network structures, networks with four compression ratios are implemented. The compression ratio is defined as the ratio of the number of elements of the intermediate variable output by the encoder to the original weight of the original CSI matrix. That is:

[0099]

[0100] Among them, numel(·) represents counting the number of elements in a matrix. For each network structure, four compression ratios are selected, and together with the proposed scheme of the present invention, experiments are carried out using data in three channel environments of LOS, NLOS, and O2I in the QuaDRiGa channel model, varying different communication overheads, and the experimental results are shown in Figure 7 in. Figure 7 It includes 9 subgraphs, divided into 3 rows and 3 columns. Each row represents a comparison between the proposed scheme of the present invention and a certain CSI feedback network based on autoencoders; each column represents a channel scenario. From the experimental results, the proposed scheme of the present invention has better performance than the three selected CSI feedback networks based on autoencoders under the same communication overhead represented by the horizontal axis, demonstrating the effective performance and good generalization of the CSI feedback system proposed by the present invention.

[0101] Embodiment 3: Thirdly, as Figure 8 shown, to achieve the above object, the present invention discloses a CSI feedback device based on important value transmission and unimportant value generation, including:

[0102] A matrix splitting module 11, configured to obtain a CSI matrix on the UE side, perform importance scoring on the elements in the CSI matrix based on an important element selection algorithm, perform descending order sorting based on the importance scoring values, and select the Top-k important elements and the index matrix of the important elements according to the sorting result;

[0103] An element encoding module 12, configured to encode the important elements using an adaptive quantization coding length algorithm to obtain a first encoding result, encode the index matrix of the important elements using an improved Huffman codebook to obtain a second encoding result, and merge the first encoding result and the second encoding result and transmit them to the BS side;

[0104] A matrix reconstruction module 13, configured to dequantize the first encoding result on the BS side to obtain a first decoding result, decode the second encoding result to obtain a second decoding result, merge the first decoding result and the second decoding result to obtain a reconstructed CSI matrix, input the reconstructed CSI matrix into a pre-established generator model, and output a CSI matrix reconstruction result.

[0105] Based on the same inventive concept, the present invention further provides a computer device, which includes: one or more processors, and a memory for storing one or more computer programs; the program includes program instructions, and the processor is configured to execute the program instructions stored in the memory. The processor may be a Central Processing Unit (CPU), or may also be other general-purpose processors, Digital Signal Processors (DSPs), Application Specific Integrated Circuits (ASICs), Field-Programmable Gate Arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, and is used to implement one or more instructions, specifically for loading and executing one or more instructions in the computer storage medium to implement the above method.

[0106] It should be further noted that, based on the same inventive concept, the present invention further provides a computer storage medium, on which a computer program is stored, and the computer program executes the above method when run by a processor. The storage medium may adopt any combination of one or more computer-readable media. The computer-readable media may be computer-readable signal media or computer-readable storage media. The computer-readable storage media may, for example, be but are not limited to electrical, magnetic, optical, electrical, magnetic, infrared, or semiconductor systems, devices, or components, or any combination of the above. More specific examples (non-exhaustive list) of the computer-readable storage media include: electrical connections with one or more wires, portable computer disks, hard disks, Random Access Memories (RAMs), Read Only Memories (ROMs), Erasable Programmable Read Only Memories (EPROMs or flash memories), optical fibers, portable compact disk read-only memories (CD-ROMs), optical storage devices, magnetic storage devices, or any suitable combination of the above. In the present invention, the computer-readable storage media may be any tangible medium that contains or stores a program, and the program may be used by or in combination with an instruction execution system, device, or component.

[0107] In the description of this specification, the descriptions referring to the terms "one embodiment", "example", "specific example", etc. mean that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present disclosure. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner.

[0108] The above has shown and described the basic principles, main features and advantages of the present disclosure. Those skilled in the art of this industry should understand that the present disclosure is not limited by the above embodiments, and the above embodiments and the descriptions in the specification only illustrate the principles of the present disclosure. Without departing from the spirit and scope of the present disclosure, the present disclosure will also have various changes and improvements, and these changes and improvements all fall within the scope of the present disclosure claimed.

Claims

1. A CSI feedback method based on important value transmission and unimportant value generation, characterized in that: The method comprises the following steps: The UE side obtains the CSI matrix, scores the importance of the elements in the CSI matrix based on the important element selection algorithm, sorts them in descending order based on the importance score values, and selects the top-k important elements and the index matrix of the important elements based on the sorting results; The important elements are encoded by using an adaptive quantization coding code length algorithm to obtain a first encoding result, the index matrix of the important elements is encoded by using an improved Huffman code book to obtain a second encoding result, and the first encoding result and the second encoding result are combined and transmitted to the BS side; The BS side dequantizes the first encoding result to obtain a first decoding result, decodes the second encoding result to obtain a second decoding result, merges the first decoding result with the second decoding result to obtain a reconstructed CSI matrix, inputs the reconstructed CSI matrix into a pre-established generator model, and outputs a CSI matrix reconstruction result.

2. A CSI feedback method based on important value transmission and unimportant value generation according to claim 1, characterized in that: The calculation process of scoring the importance of elements in the CSI matrix based on the important element selection algorithm is as follows: The importance score consists of two dimensions: the magnitude and the difference with adjacent elements; Among them, the importance score score i,j , where i and j represent the row index and column index of the element in the CSI matrix respectively, and the calculation formula is as follows: score i,j =a i,j +d(i,j),#(1) In the formula, a i,j is the amplitude, and d(i,j) is the difference between adjacent elements.

3. A CSI feedback method based on important value transmission and unimportant value generation according to claim 2, characterized in that: The amplitude a i,j The calculation formula for the difference between adjacent elements d(i,j) is as follows: Among them, h i,j is the element in the i-th row and j-th column of the CSI matrix H, real(·) represents the real part of a complex value, and imag(·) represents the imaginary part of a complex value.

4. The CSI feedback method based on important value transmission and unimportant value generation according to claim 1, characterized in that: The process of descending sorting based on importance score values: Select the top k elements with the highest scores and record their index information. Represent these k complex elements by their real and imaginary parts, and then arrange them in order of index to form an important value vector v im , for the index information of the important value vector, construct the index matrix M.

5. The CSI feedback method based on important value transmission and unimportant value generation according to claim 1, characterized in that: The process of encoding the index matrix of important elements using the improved Huffman codebook is as follows: For the compression coding of the index matrix M, an improved Huffman coding codebook C is introduced to encode the relative position of each 1 in M. First, the relative position of 1 in M ​​is counted to form a relative position vector v p , and then use C to v p Encode each element in to form a binary codeword, and then concatenate the binary codewords of all elements to form a binary string v that encodes M i ; The steps include: relative position acquisition, encoding the relative position using an improved Huffman codebook, and restoring an index matrix from the encoded binary string.

6. The CSI feedback method based on important value transmission and unimportant value generation according to claim 1, characterized in that: The process of encoding important elements using an adaptive quantization encoding code length algorithm: For the important value vector v im Quantization, encoding and decoding, first quantize the important value, then encode it, when encoding, use two lengths of code length q L and q S Encoding, where the shorter code length q S Dynamic adjustment is performed based on the important value vector to be transmitted; The steps include: quantizing the important values, determining the short code length using an adaptive method, encoding the important values, and decoding the important values.

7. The CSI feedback method based on important value transmission and unimportant value generation according to claim 1, characterized in that: The process of inputting the reconstructed CSI matrix into a pre-established generator model and outputting the CSI matrix reconstruction result: Initialize a fully connected layer and transform H m Input, multiply the output by M element by element to get the masked feature matrix F m , where the eigenvectors of non-important element positions are all zero vectors; Initialize the filled feature vector v p , copy and fill in the positions of all non-important elements, F m All the eigenvectors of the non-important elements in are represented by v p Instead, we get the filled feature matrix F p ; Use multiple layers of Transformer layers to process the feature vector sequence. For the two-dimensional matrix F composed of feature vectors p First, we cut the data by columns. The feature vector of each column is regarded as a feature vector sequence. All feature vector sequences are regarded as a batch and input into 8 series Transformer layers for processing to obtain the processed feature vector matrix F. pc , and use the same method to p Split by row, and then process to obtain the eigenvector matrix F pl , F pc and F pl Concatenate along the dimension of the feature vector to obtain the feature matrix F after the Transformer layer processing out ; Initialize a fully connected layer and input F out , output matrix H with the same dimensions as the CSI matrix out , H out The important elements in the position are replaced by H m The value of the corresponding position is obtained to obtain the final CSI reconstruction result H.

8. A CSI feedback device based on important value transmission and unimportant value generation, characterized in that: include: The matrix splitting module is used to obtain the CSI matrix on the UE side, score the importance of the elements in the CSI matrix based on the important element selection algorithm, sort them in descending order based on the importance score values, and select the Top-k important elements and the index matrix of the important elements according to the sorting results; An element encoding module is used to encode important elements using an adaptive quantization encoding code length algorithm to obtain a first encoding result, encode the index matrix of the important elements using an improved Huffman code book to obtain a second encoding result, and combine the first encoding result with the second encoding result and transmit them to the BS side; The matrix reconstruction module is used to dequantize the first encoding result on the BS side to obtain a first decoding result, decode the second encoding result to obtain a second decoding result, merge the first decoding result with the second decoding result to obtain a reconstructed CSI matrix, input the reconstructed CSI matrix into a pre-established generator model, and output the CSI matrix reconstruction result.

9. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that: The memory stores a computer program that can be run on the processor. When the processor loads and executes the computer program, a CSI feedback method based on important value transmission and unimportant value generation according to any one of claims 1 to 7 is adopted.

10. A computer-readable storage medium having a computer program stored therein, characterized in that: When the computer program is loaded and executed by the processor, a CSI feedback method based on important value transmission and unimportant value generation according to any one of claims 1 to 7 is adopted.

Images