An online-updated CSI feedback method and apparatus

Abstract

The present invention relates to the technical field of CSI feedback algorithms, and specifically provides an online updated CSI feedback method and device. Based on incremental learning, the specific steps are as follows: S1. Generate downlink CSI data; S2. Perform data cleaning; S3. Divide the data set into training data for five tasks; S4. Output a batch of training data B belonging to task c from the data stream n , and input each sample of B n and the replay subset M sub into the encoder to obtain the feature set γ n and γ m ; S5. Calculate the average similarity between each feature value in γ n and γ m ; S6. Mix the input batch B n and the replay batch B m and input them into the model for update; S7. Update the replay set of existing categories; S8. Score the samples according to the distance between each sample and the class center; S9. Repeat steps S4 to S8 until the training is completed after all the data of all tasks are input into the model. Compared with the prior art, the present invention improves the overall performance of the model on all tasks.

Description

Technical Field

[0001] The present invention relates to the technical field of CSI feedback algorithms, and specifically provides an online updated CSI feedback method and device. Background Art

[0002] In recent years, Deep Learning has shown performance comparable to or even exceeding that of humans in many fields. Researchers have found that deep learning technology not only has excellent performance in image-related fields, but also has outstanding applications in natural language processing and other aspects. Therefore, many researchers in the communication field have also begun to try to apply deep learning technology to solve communication problems and have achieved a series of excellent results. The communication algorithm based on deep learning can work in an end-to-end form, saving a large amount of manpower and material resources. At the same time, the communication algorithm based on deep learning can more effectively solve problems such as channel estimation, signal detection, and CSI feedback.

[0003] However, the above CSI feedback model based on deep learning belongs to a static model and cannot be dynamically updated over time. When the model needs to be updated, the previous training data and new data need to be input into the model for update at the same time, which poses relatively high requirements for the storage of training data and the time overhead of training.

[0004] The models running in the real world often operate in a continuous information flow environment, which requires the model to have good dynamic update capabilities while limiting the overhead of storage and training time within a certain range. How to overcome catastrophic forgetting to perform incremental training on the model is a technical problem that those skilled in the art urgently need to solve. Summary of the Invention

[0005] The present invention aims at the above deficiencies of the prior art and provides a practical online updated CSI feedback method.

[0006] A further technical task of the present invention is to provide an online updated CSI feedback device with reasonable design, safety and applicability.

[0007] The technical solution adopted by the present invention to solve its technical problems is as follows:

[0008] An online updated CSI feedback method, based on incremental learning, the specific steps are as follows:

[0009] S1. Use the 5GToolBox in Matlab software to simulate and generate downlink CSI data;

[0010] S2. Perform data cleaning. If there is a missing value in a certain CSI data, directly delete the CSI data with the missing value;

[0011] S3. Divide the dataset into training data for five tasks, process the CSI data of the five tasks into input acceptable to the model, and process the CSI data of each task into an N×2×N a ×N t , N is the number of samples, 2 represents that the channel data is divided into real and imaginary parts, N a Represents the number of rows of valid data in the original data, N t is the number of antennas of the base station;

[0012] S4. Output a batch of training data B belonging to task c from the data stream n , randomly sample the samples in the replay set M to obtain a replay subset M sub , B n and replay subset M sub Each sample in is input into the encoder to obtain the feature set γ n and γ m ;

[0013] S5. Calculation of γ n Each eigenvalue and γ in m The average similarity of training data B is calculated according to the average similarity. n Score and select the samples with the highest scores as the replay batch data B m ;

[0014] S6. Input batch B n and repeat batch B m Mix and feed into the model for updating;

[0015] S7. First, update the replay set of the existing category. Second, create a replay set for task c. n Input it into the model together with the samples in the replay set, get the output of the encoder, and calculate the class center of task c;

[0016] S8. Score the samples according to the distance between each sample and the class center, and sort the input batch B according to the score. n Sort the original samples in the replay set, select the sample with the highest score and add it to the replay set until the replay set is full of samples;

[0017] S9. Repeat steps S4 to S8 until the data of all tasks are input into the model to complete the training.

[0018] Further, in step S1, for CSI feedback data, the NRCDL channel model defined in TR38.901 of 3GPP R15 and the 5GToolBox simulation in Matlab software are used to generate downlink CSI data;

[0019] The uplink frequency of the channel is set to 2 GHz, and the downlink frequency is set to 2.1 GHz; the number of antennas of the base station is 32, and the number of antennas of the user terminal is 2.

[0020] Further, in step S3, the data set is divided into training data for five tasks: CDL-A, CDL-B, CDL-C, CDL-D, and CDL-E. The CSI data for the five tasks is processed into an input acceptable to the model, and the channel data H is transformed from the spatial-frequency domain to the angle-delay domain through discrete Fourier transform;

[0021] H′ = F c HF l H , where H is the CSI data in the spatial-frequency domain, and F c is a DFT matrix of dimension N c ×N c , and F t H is a DFT matrix of dimension N t ×N t , and H′ is the CSI data in the angle-delay domain;

[0022] Only the first N a rows of data of H′ are taken to obtain H a . H a is an N×2×N a ×N t matrix, where N is the number of samples, 2 represents that the channel data is divided into real and imaginary parts, N a represents the number of valid data rows in the original data, and N t is the number of antennas of the base station.

[0023] Further, in step S4, the training data for the five tasks of CDL-A, CDL-B, CDL-C, CDL-D, and CDL-E form a data stream, and a batch of training data B n is output in real time from the data stream. Random sampling is performed on the samples in the replay set M to obtain a replay subset M sub ;

[0024] The size of the replay subset M sub is smaller than the size of the replay set M, that is, |M sub | << |M|. Each sample in B n and M sub is input into the encoder to obtain the feature sets γ n and γ m .

[0025] Further, in step S5, the feature set γ nis a matrix with the shape of n×l, where n represents that there are n samples in an input batch, and l indicates that the output dimension of the encoder is l - dimensional;

[0026] Feature set γ m is a matrix with the shape of m×l, where m represents the replay subset M sub has m samples, and calculate the similarity matrix where β is a matrix with the shape of m×n, is the transpose matrix of γ n Calculate the average value of each row of β to obtain a similarity score vector of length m, and select several samples with the highest scores as the replay batch data B m .

[0027] Furthermore, in step S6, mix the input batch B n and the replay batch B m to obtain the combined batch B, and input B into the model for forward propagation to obtain the reconstructed CSI data B′;

[0028] Calculate the reconstruction loss L = MSE(B, B′) and the gradient, and perform backpropagation to update the model parameters.

[0029] Furthermore, in step S7, the replay set is a set of samples with a fixed size, and the size of the replay set evenly distributed to each task is also fixed. When data belonging to a new task needs to be stored in the replay set, update the replay sets of existing tasks;

[0030] Due to the addition of a new task, the size of a certain replay set becomes n′. When updating, discard several samples with lower rankings and retain the first n′ samples. When the replay set is established, it is a sequence with priorities, and the samples with higher rankings have higher priorities. Discard the samples with lower rankings to maintain the performance of the replay set;

[0031] Establish a replay set for a new category. Mix the input batch B n and the replay set M of task c c to obtain the data set BM, input BM into the model to obtain the output of the encoder, and calculate the class center of task c:

[0032]

[0033] where, E c is the class center of the new task, |BM| is the size of the data set BM, φ is the encoder, and Θ φ are the parameters of the encoder.

[0034] Furthermore, in step S8, score the samples according to the distance between each sample and the class center. The scoring logic is that if the sample xi Adding the replay set can minimize the distance between the class center of the replay set of task c and the class center E calculated from the training data, and then x c is added to the replay set: i where M

[0035]

[0036] is the replay set of task c, and the arg min function returns the x that minimizes the result of the formula in the parentheses; c i ;

[0037] Repeat this scoring process until the replay set is full.

[0038] An online updated CSI feedback device includes: at least one memory and at least one processor;

[0039] The at least one memory is used to store machine-readable programs;

[0040] The at least one processor is used to call the machine-readable program and execute an online updated CSI feedback method based on incremental learning.

[0041] Compared with the prior art, an online updated CSI feedback method and device of the present invention have the following outstanding beneficial effects:

[0042] The present invention uses a replay-based incremental learning method. First, the similarity between the input samples and the replay set samples is used to score some of the replay set samples, and the replay samples that can best help the model improve its generalization performance are selected for replay training.

[0043] Secondly, the output of the encoder of the CSI feedback model is used to calculate the class center and establish a replay set, which can establish a better replay set to help the model overcome catastrophic forgetting and improve the overall performance of the model on all tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] 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 the description of the embodiments or the prior art. Obviously, the following drawings are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0045] Attached Figure 1 is a schematic flowchart of an online updated CSI feedback method. DETAILED DESCRIPTION OF THE INVENTION

[0046] ​To enable those skilled in the art to better understand the solution of the present invention, the present invention will be further described in detail below in conjunction with specific embodiments. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the scope of protection of the present invention.

[0047] The following gives an optimal embodiment:

[0048] As Figure 1 shown, an online update CSI feedback method in this embodiment is specifically as follows:

[0049] S1. Use the 5GToolBox in Matlab software to simulate and generate downlink CSI data;

[0050] For the CSI feedback data, considering the FDD large-scale MIMO scenario, the NR CDL channel model defined in 3GPP R15 in TR38.901 and the 5GToolBox in Matlab software are used to simulate and generate downlink CSI data. The uplink frequency of the channel is set to 2 GHz, and the downlink frequency is set to 2.1 GHz. The number of antennas of the base station is 32, and the number of antennas of the user terminal is 2.

[0051] S2. Perform data cleaning. If there is a missing value in a certain CSI data, directly delete the CSI data with the missing value;

[0052] S3. According to the channel environment, divide the data set into training data for five tasks of CDL-A, CDL-B, CDL-C, CDL-D, and CDL-E. Process the CSI data of the five tasks into inputs acceptable to the model.

[0053] The channel data H is transformed from the spatial-frequency domain to the angle-delay domain through the Discrete Fourier Transform (DFT).

[0054] H′ = F c HF l H , where H is the CSI data in the spatial-frequency domain, F c is the DFT matrix of dimension N c ×N c , F t H is the DFT matrix of dimension N t ×N t , and H′ is the CSI data in the angle-delay domain. Since the first N of H′ aRows contain valid data with relatively large values, and the remaining data is basically very small or even zero. Therefore, only the first N a rows of H' are taken to obtain H a . H a is a matrix of N×2×N a ×N t . N is the number of samples, 2 represents that the channel data is divided into real and imaginary parts, and N a represents the number of rows of valid data in the original data, and N t is the number of antennas of the base station.

[0055] The training data of the five tasks of S4, CDL-A, CDL-B, CDL-C, CDL-D, and CDL-E constitutes a data stream. A batch of training data B is output in real time from the data stream n . Random sampling is performed on the samples in the replay set M to obtain a replay subset M sub . The replay subset M sub is much smaller in size than the replay set M, that is, |M sub | << |M|.

[0056] The replay subset M sub can prevent some samples in the replay set M from being retrieved repeatedly and increase the randomness of the replay samples, which can further improve the effect of replay training. Input each sample in B n and M sub into the encoder to obtain the feature sets γ n and γ m .

[0057] S5. The feature set γ n is a matrix with a shape of n×l. n represents that there are n samples in an input batch, and l indicates that the output dimension of the encoder is l-dimensional. The feature set γ m is a matrix with a shape of m×l. m represents that there are m samples in the replay subset M sub . Calculate the similarity matrix where β is a matrix with a shape of m×n, is the transpose matrix of γ n . Calculate the average value for each row of β to obtain a similarity score vector of length m. Select several samples with the highest scores as the replay batch data B m .

[0058] S6. Input batch B n and replay batch B mMix to obtain the combined batch B, and input B into the model for forward propagation to obtain the reconstructed CSI data B'. Calculate the reconstruction loss L = MSE(B, B') and the gradient, and perform backpropagation to update the model parameters.

[0059] S7. The replay set is a set of samples with a fixed size. Therefore, the size of the replay set evenly distributed to each task is also fixed. When data belonging to a new task needs to be stored in the replay set, the replay sets of existing tasks need to be updated.

[0060] Specifically, due to the addition of a new task, the size of a certain replay set becomes n'. When updating, several samples with lower rankings need to be discarded, and the first n' samples are retained. Since the replay set is a sequence with priorities when it is established, the samples with higher rankings have higher priorities. Therefore, discarding the samples with lower rankings can maximize the performance of the replay set.

[0061] Establish a replay set for the new category. Combine the input batch B n and the replay set M of task c c to obtain the data set BM. Input BM into the model to obtain the output of the encoder, and calculate the class center of task c:

[0062]

[0063] where E c is the class center of the new task, |BM| is the size of the data set BM, φ is the encoder, and Θ φ are the parameters of the encoder.

[0064] S8. Score the samples according to the distance between each sample and the class center. The scoring logic is that if adding the sample x i to the replay set can minimize the distance between the class center of the replay set of task c and the class center E c calculated from the training data, then x i is added to the replay set:

[0065]

[0066] where M c is the replay set of task c. The arg min function returns the x i when the formula in the parentheses takes the minimum value.

[0067] Repeat this scoring process until the replay set is full.

[0068] S9. Repeat steps S4 to S8 until the training is completed after all task data has been input into the model.

[0069] Based on the above method, an online updated CSI feedback device in this embodiment includes: at least one memory and at least one processor;

[0070] The at least one memory is used to store a machine-readable program;

[0071] The at least one processor is used to call the machine-readable program to execute an online updated CSI feedback method based on incremental learning.

[0072] The above-mentioned specific implementations are only specific cases of the present invention. The patent protection scope of the present invention includes but is not limited to the above-mentioned specific implementations. Any technical solutions that comply with the claims of the present invention and any appropriate changes or substitutions made by ordinary technicians in the relevant technical field shall fall within the patent protection scope of the present invention.

[0073] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An online update CSI feedback method, characterized in that, Based on incremental learning, the specific steps are as follows: S1. Use the 5GToolBox in Matlab software to simulate and generate downlink CSI data; S2. Perform data cleaning. If there are missing values in a certain CSI data, directly delete the CSI data with missing values; S3. Divide the dataset into training data for five tasks: CDL-A, CDL-B, CDL-C, CDL-D, and CDL-E. Process the CSI data of the five tasks into inputs acceptable to the model, and transform the channel data H from the spatial-frequency domain to the angle-delay domain through discrete Fourier transform; H' = F c HF l H , where H is the CSI data in the spatial-frequency domain, and F c is a DFT matrix of dimension N c ×N c , and F t H is a DFT matrix of dimension N t ×N t . Here, H' is the CSI data in the angle-delay domain; Only take the first N rows of H' a to obtain H a， H a is an N×2×N a ×N t matrix, where N is the number of samples, 2 represents that the channel data is divided into real and imaginary parts, and N a represents the number of rows of valid data in the original data, and N t is the number of antennas of the base station; The training data of five tasks, namely S4, CDL-A, CDL-B, CDL-C, CDL-D, and CDL-E, form a data stream, and a batch of training data B is output in real time from the data stream. n , randomly sample the samples in the replay set M to obtain a replay subset M. sub ; Replay subset M sub is smaller than the size of the replay set M, i.e., |M sub | << |M|. Input each sample in B n and M sub into the encoder to obtain the feature sets γ n and γ m ; S5. Calculate γ n for each eigenvalue in m and the average similarity of γ n score the training data B according to the average similarity, and select several samples with the highest scores as the replay batch data B m ; The specific operation steps are: Feature set γ n is a matrix with the shape of n×l, where n represents that there are n samples in an input batch, and l indicates that the output dimension of the encoder is l-dimensional; Feature set γ m is a matrix of shape m×l, where m represents the replay subset M sub There are m samples in it, and the similarity matrix is calculated where β is a matrix of shape m×n, is the transpose matrix of γ n For each row of β, calculate the average value to obtain a similarity score vector of length m, and select several samples with the highest scores as the replay batch data B m ; S6. Input batch B n and replay batch B m are mixed and input into the model for update; Input batch B n and replay batch B m are mixed to obtain combined batch B, and B is input into the model for forward propagation to obtain the reconstructed CSI data B'; Calculate the reconstruction loss L = MSE(B, B′) and the gradient, perform backpropagation, and update the model parameters; S7. First, update the replay sets of existing categories. Second, establish the replay set for task c, and input the input batch B n and the samples in the replay set into the model together to obtain the output of the encoder, and calculate the class center of task c; The specific operation steps are: The replay set is a set of samples with a fixed size. The size of the replay set evenly distributed to each task is also fixed. When data belonging to a new task needs to be stored in the replay set, update the replay set of the existing tasks; Due to the addition of a new task, the size of a certain replay set becomes n′. When updating, discard several samples with lower rankings and retain the first n′ samples. When the replay set is established, it is a sequence with priorities. The samples with higher rankings have higher priorities. Discard the samples with lower rankings to maintain the performance of the replay set; Create a replay set of a new category, with the input batch B n and the replay set M of task c c Mix them to obtain the dataset BM, input BM into the model, obtain the output of the encoder, and calculate the class center of task c: Among them, E c is the class center of the new task, |BM| is the size of the dataset BM, φ is the encoder, and Θ φ are the parameters of the encoder; S8. Score the samples according to the distance between each sample and the class center, and sort the samples in the input batch B n and the original samples in the replay set, select the sample with the highest score and add it to the replay set until the replay set is full of samples; S9. Repeat steps S4 to S8 until the training is completed after all task data have been input into the model.

2. The online update CSI feedback method according to claim 1, wherein In step S1, for CSI feedback data, use the NR CDL channel model defined in 3GPP R15 in TR38.901 and the 5GToolBox in Matlab software to simulate and generate downlink CSI data; The uplink frequency of the channel is set to 2 GHz, and the downlink frequency is set to 2.1 GHz; The number of antennas of the base station is 32, and the number of antennas of the user terminal is 2.

3. An online update CSI feedback method according to claim 2, characterized in that, In step S8, the samples are scored according to the distance between each sample and the class center. The scoring logic is that if the sample x i is added to the replay set, the class center of the replay set of task c and the class center E c calculated from the training data will have the minimum distance, then x i will be added to the replay set: Among them, M c is the replay set of task c, and the arg min function returns the x that minimizes the result of the formula within the parentheses i ; Repeat this scoring process until the replay set is full.

4. An online-updated CSI feedback device, characterized in that, Including: At least one memory and at least one processor; The at least one memory is used to store machine-readable programs; The at least one processor is used to call the machine-readable program and execute the method according to any one of claims 1 to 3.

Images