A channel state information compression method, system and computer device based on a sample set

Abstract

The application discloses a kind of channel state information compression method, system and computer equipment based on example set, belong to wireless communication technical field.The steps of the present application are as follows: the obtained channel state information is expressed by tensor T, and tensor, wherein tensor R represents the real part of the tensor T, and tensor I represents the imaginary part of the tensor T;Tensor R and the last two axes of tensor I are flattened, and the last axis of the tensor R and tensor I that are flattened are spliced to obtain matrix C, wherein the shape of C is [t,2*r*s];Repeat steps S1-S2 V times to obtain V matrix C;V matrix C is sequentially arranged to form matrix C list, wherein the i-th matrix in the matrix C list is represented by symbol, and the matrix C list is a channel state information example set.The present application can solve the problem of lower restoration degree under the same compression ratio, or higher compression ratio under the same restoration effect.

Description

Technical Field

[0001] This invention relates to the field of wireless communication technology, and more specifically, to a method, system, and computer device for channel state information compression based on a sample set. Background Technology

[0002] With the development of wireless communication technology, especially MIMO technology, the number of antennas at the signal transmitting and receiving ends is increasing, leading to a significant increase in the size of the channel state information tensor describing the channel state information. Consequently, the transmission cost of channel state information is also increasing. To reduce the transmission cost of channel state information, the industry has developed schemes that use neural networks to compress and reconstruct the channel state information tensor, such as invention CN201811488677.9. Although these schemes have achieved good results, there is still considerable room for improvement. At the same compression ratio, the restoration accuracy is low, or at the same restoration effect, the compression ratio is high (resulting in a larger compressed file). Summary of the Invention

[0003] The present invention aims to overcome at least one of the defects of the prior art and provide a channel state information compression method, system and computer device based on sample set, which is used to solve the problem of low restoration degree under the same compression ratio, or high compression ratio under the same restoration effect, and further improve the compression and reconstruction effect of channel state information tensor.

[0004] According to one aspect of the present invention, a method for channel state information compression based on a sample set is provided. This method for channel state information compression based on a sample set includes the following steps:

[0005] F1. Construct a sample set C of channel state information;

[0006] F2. Receive the channel state information X to be compressed;

[0007] F3. Calculate the similarity matrix C of X. * ;

[0008] F4. Calculate C * The position pos in C;

[0009] F5. Through X and C * Obtain the optimal scalar ;

[0010] F6. Through X and C * and optimal scalar Construct the input parameter RES;

[0011] F7. Construct a compressor to compress RES and generate a vector. ;

[0012] F8. For vectors Quantization is performed to generate a binary vector. ;

[0013] F9. Output about ,pos, binary vector .

[0014] To complete a painting, starting from scratch would require a great deal of effort. However, if a draft of the painting is provided beforehand, and the painting is completed by refining details from the draft, it is obviously much easier. Similarly, to reconstruct channel state information from scratch, a large amount of information cues (the compressed channel state information) are needed. However, if the approximate structure of the channel state information is provided, only a small amount of information cues are needed to accurately reconstruct the original channel, thus improving the compression effect. The solution of this invention constructs a sample set of channel state information tensors, and then calculates the nearest channel state information tensor in the sample set to the compressed channel state information tensor as the "approximate structure" of the compressed channel state information tensor. This is used to improve the model's reconstruction effect under the same compression ratio.

[0015] Step F1, constructing the channel state information sample set C, specifically involves:

[0016] S1. Represent the obtained channel state information using a tensor T, wherein the tensor T has the shape [t, r, s], where t is the delay spread, r is the number of receive antennas, and s is the number of transmit antennas, and the tensor... , where tensor R represents the real part of tensor T, and tensor I represents the imaginary part of tensor T;

[0017] S2. Flatten the last two axes of tensor R and tensor I, and concatenate the flattened tensor R with the last axis of tensor I to obtain matrix C, where the shape of C is [t, 2*r*s].

[0018] S3. For the obtained V state information of the channel, transform them according to the method of steps S1-S2 to obtain V different matrices C;

[0019] S4. Arrange the V distinct matrices C in order to form a list of matrices C, where the i-th matrix in the list of matrices C is denoted by the symbol... This indicates that the matrix C list is a set of channel state information samples.

[0020] The channel information is measured in real-world scenarios, and a channel state information tensor T is generated. By using a large amount of channel information obtained from real-world scenarios, the approximate structure of the channel state information is simulated. Only a small amount of information is needed to accurately reconstruct the original channel.

[0021] Step F3. Calculate the similarity matrix C of X. * Specifically:

[0022] ,

[0023] in, , .

[0024] The distance function LS_loss is used to find the most similar matrix from a set of matrices.

[0025] Step F5. Passing through X and C * Obtain the optimal scalar Specifically:

[0026]

[0027] in, [m,n] represents X, The shape.

[0028] Step F6. Passing through X and C * and optimal scalar The input parameter RES is constructed as follows:

[0029] RES= .

[0030] Step F8. For vectors Quantization is performed to generate a binary vector. Specifically:

[0031] ,

[0032] in, , For the updated moving average vector, for The i-th element, The integer symbol.

[0033] Applying quantization to channel generation can significantly reduce the size of the latent variable's value space, resulting in better model generation performance.

[0034] The update method is as follows:

[0035] ,

[0036] in, It is a hyperparameter that is manually set and ranges from 0 to 1.

[0037] Step F9. Output about ,pos, binary vector Specifically:

[0038] The vector The first to the a-th elements construct a float a format. The (a+1)th to the 2ath elements construct a float a format. The 2a+1th to the 2a+1th The element constructs the location information pos, the first element... One to the first Construct a binary vector from each element. , where d is a binary vector Dimensions.

[0039] forward The first d elements record the "general structure" of the channel state information, while the last d elements record supplementary information. The cost of compression in this invention is... This significantly reduces the number of storage locations (d) required to achieve the same compression rate, thus greatly reducing the size of the storage location (d) needed.

[0040] According to another aspect of the present invention, a channel state information compression system based on a sample set is provided. This channel state information compression system based on a sample set includes:

[0041] The channel state information sample set construction module is used to construct the channel state information sample set C;

[0042] The channel receiving module is used to receive the channel state information X to be compressed;

[0043] A compression module is used to calculate the similarity matrix C of X. * ; Calculate C * The position pos in the list of matrix C; through X and C * Obtain the optimal scalar ; via X and C * and optimal scalar Construct the input parameters RES; use a neural network to build a compressor to compress RES and generate a vector. ;

[0044] Output module, used for vectors Quantization is performed to generate a binary vector. Output about ,pos, binary vector .

[0045] According to another aspect of the present invention, a computer device is provided. The computer device includes a builder and a processor, the builder having constructed a computer program, and the processor, when executing the computer program, implementing a channel state information compression method based on a sample set.

[0046] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0047] This invention proposes a channel state information compression method and system based on a sample set. By reducing the size of the value space of latent variables, the model generation effect is improved. Through quantization, sorting compression (preserving important delay spreads according to the standard deviation), and allocating the storage space of the latent vector according to the importance of the delay spread, the value space of the latent variables is reduced. This is used to solve the problem of low restoration degree under the same compression ratio, or high compression ratio under the same restoration effect, and further improves the compression and reconstruction effect of channel state information tensor. Attached Figure Description

[0048] Figure 1 This is a flowchart of a channel state information compression method based on a sample set according to the present invention.

[0049] Figure 2 This is a system block diagram of a channel state information compression system based on a sample set according to the present invention. Detailed Implementation

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention are within the scope of protection of the present invention.

[0051] Example 1

[0052] like Figure 1 As shown, according to an embodiment of the present invention, a channel state information compression method based on a sample set is provided. This channel state information compression method based on a sample set includes the following steps:

[0053] F1. Construct a sample set C of channel state information;

[0054] F2. Receive the channel state information X to be compressed;

[0055] F3. Calculate the similarity matrix C of X. * ;

[0056] F4. Calculate C * The position pos in C;

[0057] F5. Through X and C * Obtain the optimal scalar ;

[0058] F6. Through X and C * and optimal scalar Construct the input parameter RES;

[0059] F7. Construct a compressor to compress RES and generate a vector. ;

[0060] F8. For vectors Quantization is performed to generate a binary vector. ;

[0061] F9. Output about ,pos, binary vector .

[0062] Specifically, a compressor is built using a neural network.

[0063] Step F1, constructing the channel state information sample set C, specifically involves:

[0064] S1. Represent the obtained channel state information using a tensor T, wherein the tensor T has the shape [t, r, s], where t is the delay spread, r is the number of receive antennas, and s is the number of transmit antennas, and the tensor... , where tensor R represents the real part of tensor T, and tensor I represents the imaginary part of tensor T;

[0065] S2. Flatten the last two axes of tensor R and tensor I, and concatenate the flattened tensor R with the last axis of tensor I to obtain matrix C, where the shape of C is [t, 2*r*s].

[0066] S3. For the obtained V state information of the channel, transform them according to the method of steps S1-S2 to obtain V different matrices C;

[0067] S4. Arrange the V distinct matrices C in order to form a list of matrices C, where the i-th matrix in the list of matrices C is denoted by the symbol... This indicates that the matrix C list is a set of channel state information samples.

[0068] In real-world scenarios, measure and generate channel information, and generate a channel state information tensor.

[0069] Step F3. Calculate the similarity matrix C of X. * Specifically:

[0070] ,

[0071] in, , .

[0072] Specifically, It refers to the value of the variable that makes the following expression reach its minimum value.

[0073] Step F5. Passing through X and C * Obtain the optimal scalar Specifically:

[0074]

[0075] in, [m,n] represents X, The shape.

[0076] Specifically, The function is used to calculate the distance between two points.

[0077] Step F6. Passing through X and C * and optimal scalar The input parameter RES is constructed as follows:

[0078] RES= .

[0079] For convenience, the input is represented by the matrix RES.

[0080] Step F8. For vectors Quantization is performed to generate a binary vector. Specifically:

[0081] ,

[0082] in, , For the updated moving average vector, for The i-th element, The integer symbol.

[0083] Specifically, c = 0.5. The Sigmoid function is a common sigmoid function in biology, also known as an sigmoid growth curve. In information science, due to its monotonically increasing and inversely monotonically increasing properties, the Sigmoid function is often used as an activation function for neural networks, mapping variables to the range of 0 and 1.

[0084] The update method is as follows:

[0085] ,

[0086] in, It is a parameter ranging from 0 to 1.

[0087] Specifically, It is a hyperparameter that is manually set and ranges from 0 to 1.

[0088] Step F9. Output about ,pos, binary vector Specifically:

[0089] The vector The first to the a-th elements construct a float a format. The (a+1)th to the 2ath elements construct a float a format. The 2a+1th to the 2a+1th The element constructs the location information pos, the first element... One to the first Construct a binary vector from each element. , where d is a binary vector Dimensions.

[0090] Specifically, a=16. During training, a typical VAE approach ensures that all elements of the latent variables follow a certain distribution by adding a KL divergence loss term to the loss function. Then, during generation, the elements of the latent variables are randomly sampled from this distribution and used as input to the decoder for the generation task. This process relies on the assumption that the sampling distributions of the latent variables during training and generation are the same. However, since the elements of the latent variables are correlated during training, this assumption does not hold true in most scenarios. Reducing the value space of the latent variables can alleviate the inconsistency between the latent variable distributions during training and generation. Therefore, this invention proposes to improve the model's generation performance by reducing the size of the latent variable value space.

[0091] Example 2

[0092] like Figure 2 As shown, according to an embodiment of the present invention, a channel state information compression system based on a sample set is provided. This channel state information compression system based on a sample set includes:

[0093] Channel state information sample set construction module 210 is used to construct a channel state information sample set C;

[0094] The channel receiving module 220 is used to receive the channel state information X to be compressed;

[0095] Compression module 230 is used to calculate the similarity matrix C of X. * ; Calculate C * The position pos in the list of matrix C; through X and C * Obtain the optimal scalar ; via X and C * and optimal scalar Construct the input parameters RES; use a neural network to build a compressor to compress RES and generate a vector. ;

[0096] Output module 240 is used for vectors Quantization is performed to generate a binary vector. Output about ,pos, binary vector .

[0097] Example 3

[0098] According to an embodiment of the present invention, a computer device is provided. The computer device includes a builder and a processor, the builder constructing a computer program, and the processor executing the computer program implementing a channel state information compression method based on a sample set.

[0099] Example 4

[0100] The channel state data has 126 extension values, 4 receive antennas, and 32 transmit antennas, so it can be constructed using a complex tensor T of shape [126, 4, 32]. Flattening the last two axes of the real and imaginary tensors of T and then concatenating them along the last axis yields a matrix, denoted as C, with shape [t, 2*r*s]. 10,000 examples of matrix state information were collected and constructed using matrix C, resulting in a list of matrices C of length 1000.

[0101] Let matrix X represent the channel state information to be compressed. The shape of matrix X is [128, 2*4*32].

[0102] Calculate the most similar C matrix and find The position of the scalar in the list of matrix C is denoted as pos. The two optimal scalars are then calculated.

[0103] Input parameter matrix RES.

[0104] Following the approach of using attention mechanisms in natural language processing, the amount of latency spread in X is treated as the text length, and multiple attention layers are used to process the data. The specific scheme is as follows:

[0105] The dimension of matrix X is reduced from 2*4*32 to 64 dimensions using trainable fully connected layers.

[0106] Initialize a trainable position matrix P, with shape [126, 64]. The P matrix is ​​used to enable multiple attention layers to perceive positional information.

[0107] The reduced X matrix is ​​added to the P matrix and then input into multiple attention layers. The resulting output is denoted as M, and the shape of M is [126, 64].

[0108] Flattening the M matrix yields a vector of dimension 126*64, and then using a trainable fully connected layer to reduce this vector to 454 dimensions. The resulting 454-dimensional vector is then used... express.

[0109] Update the moving average vector Based on the above description, for each matrix X, a vector can be computed. During training, vectors are used. As The estimate of the mean The update method is as follows: .

[0110] Quantization: Generating binary vectors Its dimension is 454, and it is consistent with... It has the following relationship:

[0111] Through the calculations of the compression module, the most similar matrix can be calculated for matrix X. , and its two corresponding scalars as well as The position pos in the list of matrix C. A binary vector was designed. This will be used as the final output of the compression module. The specific structure is as follows:

[0112] vector Elements 1 through 16 are stored in float16 format. .

[0113] vector Elements 17 through 32 are stored in float16 format. .

[0114] vector The 33rd to the 32+14th (because the list length of matrix C is 10000, so use...) The capacity of a bit can store index information. The element stores the location information pos.

[0115] vector The 47th to the Each element stores a binary vector. .

[0116] Obviously, binary vectors The dimension is 500.

[0117] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the technical solution of the present invention, and are not intended to limit the specific implementation of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the claims of the present invention should be included within the protection scope of the claims of the present invention.

Claims

1. A channel state information compression method based on a sample set, characterized in that, Includes the following steps: F1. Construct a sample set C of channel state information; F2. Receive the channel state information X to be compressed; F3. Calculate the similarity matrix of X. ; F4. Calculation The position pos in C; F5. Through X and Obtain the optimal scalar ; F6. Through X and and optimal scalar Construct the input parameter RES; F7. Construct a compressor to compress RES and generate a vector. ; F8. For vectors Quantization is performed to generate a binary vector. ; F9. Output about ,pos, binary vector ; Step F1, constructing the channel state information sample set C, specifically involves: S1. Represent the obtained channel state information using a tensor T, wherein the tensor T has the shape [t, r, s], where t is the delay spread, r is the number of receive antennas, and s is the number of transmit antennas, and the tensor... , where tensor R represents the real part of tensor T; S2. Flatten the last two axes of tensor R and tensor I, and concatenate the flattened tensor R with the last axis of tensor I to obtain matrix C, where the shape of C is... ; S3. For the obtained V state information of the channel, transform them according to the method of steps S1-S2 to obtain V different matrices C; S4. Arrange the V distinct matrices C in order to form a list of matrices C, where the i-th matrix in the list of matrices C is denoted by the symbol... This indicates that the matrix C list is a set of channel state information samples.

2. A channel state information compression system based on a sample set, applied to the channel state information compression method based on a sample set as described in claim 1, characterized in that, Also includes: The channel state information sample set construction module is used to construct the channel state information sample set C; The channel receiving module is used to receive the channel state information X to be compressed; A compression module used to calculate the similarity matrix of X. ;calculate The position pos in the list of matrix C; via X and Obtain the optimal scalar ; through X and and optimal scalar Construct the input parameters RES; use a neural network to build a compressor to compress RES and generate a vector. ; Output module, used for vectors Quantization is performed to generate a binary vector. ; Output about ,pos, binary vector .

3. A computer device comprising a builder and a processor, the builder comprising a computer program, characterized in that, When the processor executes the computer program, it implements the channel state information compression method based on a sample set as described in claim 1.

Images