A channel state information feedback method for ultra-large-scale MIMO

Abstract

The present invention relates to a channel state information feedback method for ultra-large-scale MIMO. The method adopts two feedback paradigms. The first feedback paradigm consists of a codec, and the second feedback paradigm consists of a sampling strategy and a decoder. The two feedback paradigms compress, feedback, and decompress the original CSI matrix in parallel, and simultaneously obtain an estimated value of the original CSI. The two feedback paradigms are trained separately using an end-to-end training method, and the trained network parameters are saved. After training, a feedback paradigm is selected. Within the paradigm selection time, the two feedback paradigms are executed successively and the downlink quality of the corresponding time period is measured. The method with better performance is selected as the subsequent feedback method. The method continuously monitors the changes in downlink quality. If the downlink quality decreases significantly, the feedback paradigm is reselected; otherwise, the feedback paradigm remains unchanged. Compared with the existing technology, the present invention enhances scalability and applicability by processing the two feedback paradigms in parallel and executing them preferentially.

Description

Technical Field

[0001] The present invention relates to the field of wireless communication technology, and in particular to a channel state information feedback method for ultra-large-scale MIMO. Background Art

[0002] As society undergoes dramatic changes in how it creates, shares, and consumes information, wireless data traffic is increasing dramatically, further fueling the demand for communication transmission speeds and capacity. Terahertz communication technology and ultra-large-scale MIMO are considered foundational enabling technologies for 6G communication systems. They achieve terahertz-scale transmission speeds by integrating a large number of nanoantenna arrays within a package and leveraging the gain generated by the antenna arrays to overcome the limitations of terahertz frequency losses and limited transmitter power. Compared to massive MIMO, ultra-large-scale MIMO systems utilize larger antenna arrays and operate at higher terahertz frequencies, enabling higher data transmission rates and greater channel capacity.

[0003] Many of the performance benefits of ultra-large-scale MIMO, such as beamforming gain, rely on obtaining complete and accurate CSI. In frequency-division multiplexing (FDM) systems, the uplink and downlink operate on different frequencies. To obtain downlink CSI, the base station first sends a pilot signal to the user end. The user end then performs channel estimation to obtain CSI and then feeds the CSI back to the base station via the uplink. In ultra-large-scale MIMO systems, the number of antennas is far greater than in large-scale MIMO systems. The resulting high-dimensional CSI matrix results in significant feedback overhead. Therefore, to ensure the integrity and accuracy of CSI matrix feedback, compression of the high-dimensional CSI matrix is ​​necessary.

[0004] Traditional CSI feedback techniques primarily rely on codebooks and compressed sensing. However, as antenna array size increases, the codebook complexity required by the former also increases, while the matrix sparsity required by the latter becomes increasingly difficult to satisfy. Consequently, traditional CSI feedback techniques struggle to meet the CSI feedback requirements of large-scale MIMO systems, let alone ultra-large-scale MIMO systems. In large-scale MIMO systems, some work has applied deep learning to address CSI feedback. These networks, such as CsiNet and CRNet, have achieved superior feedback performance compared to traditional techniques. Therefore, deep learning is particularly well-suited for CSI feedback in large-scale and even ultra-large-scale MIMO systems.

[0005] Compared to massive MIMO systems, ultra-large-scale MIMO systems employ a subarray structure and operate in the terahertz frequency band. These systems are subject to channel characteristics such as molecular absorption loss and three-dimensional spherical wave propagation. Consequently, their downlink CSI matrices are characterized by high dimensionality, large number, and small amplitude. To address the CSI feedback issue, appropriate sampling strategies and deep learning networks can exploit the frequency correlation between subcarriers and the correlation between the real and imaginary parts of the matrix to extract features for compression and reconstruction, effectively reducing feedback overhead. However, currently, available datasets and feasible feedback schemes for high-dimensional CSI feedback in ultra-large-scale MIMO scenarios remain lacking. Therefore, addressing CSI feedback in ultra-large-scale MIMO systems is an important research topic. Summary of the Invention

[0006] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and provide a channel state information feedback method for ultra-large-scale MIMO, so as to improve adaptability while reducing network training parameters and improving training effects.

[0007] The purpose of the present invention can be achieved by the following technical solutions:

[0008] The present invention provides a channel state information feedback method for ultra-large-scale MIMO, which is applied to a base station and includes the following steps:

[0009] Sending a pilot signal to a user terminal, receiving compressed first information based on a first feedback paradigm in a first time period, and receiving second information based on a second feedback paradigm obtained by sampling complete CSI corresponding to a subcarrier frequency in a second time period;

[0010] Decompressing the first information and the second information respectively, obtaining CSI matrix estimates of the ultra-large-scale MIMO in the first time period and the second time period, and based on downlink quality in the first time period and the second time period, performing CSI matrix feedback using a feedback paradigm with better downlink quality in subsequent communication processes;

[0011] The compression of the first information is achieved by an encoder that is set at the user end and pre-trained based on the original CSI matrix data set, the second information is obtained based on the sampling strategy of the user end, and the decompression is achieved by a decoder that is set at the base station end and pre-trained based on the original CSI matrix data set, and the encoder and decoder of the first feedback paradigm share the same network training parameters.

[0012] As a preferred technical solution, the original CSI matrix data set is divided into a real data set and an imaginary data set. The original CSI matrix data set is divided into a real data set and an imaginary data set. The real data set is used for training and verification, and the real data set and the imaginary data set are used for testing at the same time.

[0013] As a preferred technical solution, the two feedback paradigms are trained with a fixed or variable learning rate based on an optimizer using a mean square error function as the loss function.

[0014] As a preferred technical solution, the training process of the encoder and the decoder includes the following steps:

[0015] To minimize the difference between the original CSI matrix and the estimated CSI matrix output by the decoder, an end-to-end training method is used to train the two feedback paradigms respectively, and the feedback network parameters trained under the first feedback paradigm and the second feedback paradigm are saved.

[0016] As a preferred technical solution, the training of the two feedback paradigms is specifically as follows:

[0017] The encoder and decoder are jointly trained for the first feedback paradigm, while only the decoder is trained for the second paradigm.

[0018] As a preferred technical solution, the encoder and decoder are implemented based on a long short-term memory recursive neural network.

[0019] As a preferred technical solution, the acquisition of the original CSI matrix data set includes the following steps:

[0020] A very large-scale MIMO channel model is constructed, and fixed parameters except the Euler angle at the center of the transmit antenna array are set. Multiple CSI matrices are generated by changing the Euler angle at the center of the transmit antenna array to form the original CSI matrix data set.

[0021] As a preferred technical solution, the fixed parameters include the number of transmitting and receiving antenna subarrays, the number of antenna elements in the transmitting and receiving antenna subarrays, the three-dimensional coordinates of the centers of the transmitting and receiving antenna arrays, the Euler angles of the center of the receiving antenna array, the transmitting frequency, the bandwidth, the number of carriers, and the number of subbands in each subcarrier.

[0022] As a preferred technical solution, the channel state information feedback method further includes:

[0023] The downlink quality is detected at fixed intervals. If the quality drops by more than a certain threshold, a better feedback mode is selected for CSI matrix feedback.

[0024] As a preferred technical solution, the first feedback paradigm and the second feedback paradigm are set in parallel, and the base station is provided with two decoders with the same structure for decompressing the first information and the second information respectively.

[0025] Compared with the prior art, the present invention has the following advantages:

[0026] (1) Adopting multiple feedback paradigms, with strong adaptability and few network parameters: The present invention is based on the CSI matrix estimation values ​​of the first feedback paradigm and the second feedback paradigm corresponding to the first time period and the second time period, and adopts the feedback paradigm with better downlink quality for CSI matrix feedback in the subsequent communication process. It can select a better feedback paradigm to estimate CSI in different environments. In addition, the codec and sampling strategy can be adjusted according to the specific application. The method has strong scalability and wide adaptability. On this basis, in order to overcome the problem of large amount of training parameters brought by the two sets of feedback paradigms, the first feedback paradigm codec of the present application shares the same network training parameters, and the second feedback paradigm uses a sampling strategy to extract the complete CSI corresponding to some subcarrier frequencies for feedback. At the base station end, the correlation between the subcarriers of the ultra-large-scale MIMO channel is used to predict the estimated value of the original CSI matrix. No encoder is required. Both paradigms can effectively reduce network training parameters and are simple and easy to implement.

[0027] (2) Fast training speed: During the training process, the complex CSI matrix is ​​divided into real and imaginary matrices. The correlation between the real and imaginary numbers in the CSI matrix of ultra-large-scale MIMO is utilized to train only the real part. The real and imaginary parts share the same training network for compression and feedback, reducing the network computational scale by nearly an order of magnitude. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 Schematic diagram of the algorithm logic flow in the embodiment;

[0029] Figure 2 Schematic diagram of the first feedback paradigm in the embodiment;

[0030] Figure 3 Schematic diagram of the second feedback paradigm in the embodiment. DETAILED DESCRIPTION

[0031] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0032] Example 1

[0033] To address the problem of high CSI feedback overhead in ultra-large-scale MIMO systems, this embodiment provides a channel state information (CSI) algorithm based on sampling strategy and deep learning in ultra-large-scale multiple input multiple output (MIMO) systems. The method establishes a very large-scale MIMO channel model in the terahertz frequency band and cyclically updates the angle parameters to generate the original CSI matrix at the base station. The user end obtains the original CSI matrix through channel estimation and divides it into a real matrix and an imaginary matrix for subsequent training and verification respectively. Two feedback paradigms are simultaneously constructed. The first feedback paradigm consists of a codec, and the second feedback paradigm consists of a sampling strategy and a decoder. The two feedback paradigms compress, feedback and decompress the original CSI matrix in parallel, and obtain the estimated value of the original CSI at the same time. The two feedback paradigms are trained separately in an end-to-end training manner, and the trained network parameters are saved. After training, the feedback paradigm is selected. The two feedback paradigms are executed successively within the paradigm selection time and the downlink quality of the corresponding time period is measured. The one with better performance is selected as the subsequent feedback method. The downlink quality changes are continuously monitored. If the downlink quality deteriorates significantly, the feedback paradigm is reselected; otherwise, the feedback paradigm remains unchanged. The present invention fills the gap in CSI feedback methods in ultra-large-scale MIMO scenarios, enhances scalability and applicability by processing two feedback paradigms in parallel and executing them preferentially, and utilizes the subcarrier correlation and the real and imaginary part correlation of ultra-large-scale MIMO channels to reduce network training parameters and calculation scale.

[0034] See also Figure 1 , this method comprises the following steps:

[0035] Step 1: Establish a three-dimensional ultra-large-scale MIMO channel model and obtain the CSI matrix representation of the downlink in the ultra-large-scale MIMO scenario.

[0036] Step 2: Deploy a very large-scale MIMO antenna array based on the model established in the first step. Set fixed parameters, change the Euler angle at the center of the transmitting antenna array, and update the angle parameters to change the departure angle of the antenna subarray. Substitute the fixed parameters and angle parameters into the model to generate the CSI matrix, forming the original CSI matrix dataset at the base station.

[0037] The fixed parameters specifically include: the number of transmitting and receiving antenna subarrays, the number of antenna elements in the transmitting and receiving antenna subarrays, the three-dimensional coordinates of the centers of the transmitting and receiving antenna arrays, the Euler angle of the center of the receiving antenna array, the transmitting frequency, the bandwidth, the number of subcarriers and the number of subbands in each subcarrier; the variable parameters specifically include: the Euler angle of the center of the transmitting antenna array, the eular TX =[x,0,0] T , x is a random number that obeys a uniform distribution in the interval (-π,π). Changing the variable parameters generates multiple matrices as data sets.

[0038] Step 3: The base station sends a pilot signal to the user end. The user end receives the pilot signal sent by the base station and performs channel estimation based on the signal to obtain the original CSI matrix. The original CSI matrix is ​​the matrix in the original CSI matrix data set in step 2.

[0039] In this step, it is assumed that the channel estimation at the user end is ideal and a complete and accurate original CSI matrix for the downlink can be obtained.

[0040] Step 4: The user end divides the real and imaginary data of the original CSI matrix to obtain a real data set and an imaginary data set, which are used for training and verification respectively in step 7.

[0041] Step 5: Construct two feedback paradigms on the user side. The first paradigm is to build an encoder to convert the original CSI matrix into a compressed matrix through the encoder. The second paradigm is to build a sampling strategy to sample the original CSI matrix to obtain a sampled matrix.

[0042] Step 6: Build a decoder for the CSI feedback network at the base station. In both paradigms, the compressed matrix fed back to the base station via the uplink from the user end needs to be input into the decoder for decompression to obtain an estimate of the original CSI matrix.

[0043] Step 7: Use an end-to-end training method to train the two paradigms separately. The first paradigm jointly trains the encoder and decoder, and the second paradigm only trains the decoder. The goal is to minimize the difference between the original CSI matrix and the matrix estimate output by the decoder, and preserve the feedback network parameters trained under the two paradigms.

[0044] Among them, the loss function of the feedback network training of both feedback paradigms adopts the mean square error function, the optimizer is used to optimize the network according to the loss function, and a fixed or variable learning rate is used. The first feedback paradigm is initialized in a conventional way, and the second feedback paradigm is designed with an initialization method based on the corresponding sampling strategy, with the purpose of performing a zero-padding operation on it and initializing it to a matrix with the same dimension as the original CSI and a value of 0 at the non-sampling position.

[0045] Step 8: Select a feedback paradigm for the two trained paradigms. Set the paradigm selection time. In the first half of the time, the user end executes the first feedback paradigm, feeds back the CSI matrix to the base station, and then measures and saves the downlink quality of this period. In the second half of the time, the second feedback paradigm is executed. After the feedback, the downlink quality of this period is measured and saved. The downlink quality of the two periods is compared, and the paradigm with better performance is determined as the subsequent CSI feedback method.

[0046] Step 9: Set the interval time and measure the downlink quality of the user end at each interval. If the quality drops significantly, reselect the feedback paradigm; otherwise, the feedback paradigm remains unchanged.

[0047] Among them, the CSI matrix of the established ultra-large-scale MIMO system is a four-dimensional complex matrix where N t is the total number of antenna elements in the transmitting antenna array, N r is the total number of antenna elements in the receiving antenna array, N c is the number of subcarriers, is the number of subbands in each subcarrier.

[0048] The following is a specific implementation process of this method, which includes the following steps:

[0049] Step 1: Establish the three-dimensional ultra-large-scale MIMO channel matrix H UM , H UM is a complex matrix, represented by

[0050]

[0051] In the formula, each independent sub-channel Represents the qth t The qth transmitting subarray and the r Since the path loss in the terahertz band is very serious and the multipath component is almost negligible, it is assumed that only the line-of-sight path exists. It consists of three parts: path loss under line-of-sight path, antenna gain and equivalent antenna array response, namely

[0052]

[0053] Where, α LoS is the path loss under the line-of-sight path, G (t) and G (r) are the antenna gains of the transmitting and receiving ends respectively, A (t) (·) and A (r) (·) are the equivalent antenna array responses at the transmitting and receiving ends respectively, Φ (t) =[φ(t) ,θ (t) ] and Φ (r) =[φ (r) ,θ (r) ] are departure angle and arrival angle vectors respectively, and are the beamforming angle vectors at the transmitting and receiving ends, respectively, which are set to the line-of-sight direction from the center of the receiving antenna array to the center of the transmitting antenna array, φ is the azimuth angle, and θ is the elevation angle.

[0054] α LoS It includes two parts: propagation loss and molecular absorption loss, and is calculated as follows:

[0055]

[0056] Where c0 is the speed of light, d is the distance between the transmitting subarray and the receiving subarray, which is determined by the subarray position, and f k is the frequency of the kth subcarrier, λ is the path loss exponent, which is set to 2, and K(·) is the molecular absorption coefficient, which is obtained by using a high-resolution transmission molecular absorption database.

[0057] G (t) and G (r) It is set to 1dB in the algorithm.

[0058] A(Φ,Φ0) is calculated using the traditional uniform planar array method and is determined by the antenna subarray size, the coordinates of the antenna elements in the subarray, the departure angle / arrival angle of the subarray, and the subcarrier frequency.

[0059] The final calculated four-dimensional complex matrix It can be considered as the CSI matrix expression of the downlink in the ultra-large-scale MIMO scenario, where N t is the total number of antenna elements in the transmitting antenna array, N r is the total number of antenna elements in the receiving antenna array, N c is the number of subcarriers, is the number of subbands in each subcarrier.

[0060] Step 2: Deploy the ultra-large-scale MIMO antenna array based on the model established in Step 1 and set fixed parameters, including the number of transmit antenna subarrays Q t , the number of receiving antenna subarrays Q r , the number of antenna elements in the transmitting antenna subarray Q at , the number of antenna elements in the receiving antenna subarray Q ar , the three-dimensional geometric coordinate position of the center of the transmitting antenna array TX , the three-dimensional geometric coordinate position of the center of the receiving antenna array RX, the Euler angle of the center of the receiving antenna array RX , the transmitting frequency f c , bandwidth B, number of subcarriers N c and the number of subbands in each subcarrier The total number of transmitting antenna array antenna elements is N t =Q t ×Q at , the total number of antenna elements in the receiving antenna array is N r =Q r ×Q ar The fixed parameter values ​​in this embodiment are shown in Table 1.

[0061] Table 1 Fixed parameter setting table

[0062]

[0063]

[0064] Cyclic generation of random numbers x that obey uniform distribution in the interval (-π, π), the center Euler angle of the transmitting antenna array eular TX =[x,0,0] T , update the angle parameters, thereby changing the departure angle of the antenna subarray, and substitute the fixed parameters set in Table 1 and the updated angle parameters into the channel model established in step 1 to generate the CSI matrix H UM (1024, 1, 4, 256), and 15,000 CSI matrices are obtained at the end of the loop, forming the original CSI matrix data set at the base station.

[0065] Step 3: The base station sends a pilot signal to the user end. The user end receives the pilot signal sent by the base station and performs channel estimation based on the signal to obtain the original CSI matrix. Assuming that the channel estimation on the user end is ideal, a complete and accurate CSI matrix for the downlink can be obtained.

[0066] Step 4: The user end divides the real and imaginary data of the original CSI matrix to obtain a real data set and an imaginary data set, which are used as a training set and a validation set for training and validation in step 7, respectively.

[0067] Step 5: Construct two feedback paradigms on the user side. The first paradigm is to build an encoder to convert the original CSI matrix into a codeword vector through the encoder. The second paradigm is to build a sampling strategy to sample the original CSI matrix to obtain a sampled matrix.

[0068] In this embodiment, the first feedback paradigm is designed as follows: Figure 2As shown in Figure 1, the encoder includes a matrix reconstruction operation, a single-layer long short-term memory (LSTM) recurrent neural network, and a fully connected layer. In order to adapt to the processing mechanism of the LSTM neural network, the input four-dimensional CSI matrix H is first converted to UM (1024,1,4,256) reconstructed into a two-dimensional matrix The first dimension represents frequency, and the second dimension represents the number of antenna elements. The reconstructed matrix is ​​input into a single-layer LSTM neural network for feature extraction, and the output feature map is 1024 × 1024. This feature map is then compressed into a fully connected layer containing N neurons, where N is calculated based on the compression ratio (CR), i.e., N = 1024 × CR. The output matrix is ​​1024 × N, which is the compressed matrix that needs to be fed back to the base station.

[0069] In this embodiment, the second feedback paradigm is designed as follows: Figure 3 As shown, the specific implementation of the sampling strategy is: reconstruct the input CSI matrix into a two-dimensional matrix In the frequency dimension, sampling is performed at equal intervals according to the compression ratio CR, that is, one row is sampled for every CR row to form a sampled matrix, which is used as the compression matrix to be fed back to the base station.

[0070] Step 6: Construct a decoder for the CSI feedback network at the base station. Under both feedback paradigms, the compressed matrix fed back to the base station via the uplink from the user end needs to be input into the decoder for decompression to obtain an estimate of the original CSI matrix.

[0071] like Figure 2 and Figure 3 As shown in the figure, the decoder structure of the two feedback paradigms in this embodiment is the same, including a fully connected layer, a single-layer LSTM neural network and a matrix reconstruction operation. The compressed matrix fed back by the user end is input into the fully connected layer, and the output matrix is ​​a 1024×1024 matrix. It is then input into the single-layer LSTM neural network to extract features, and the output is a 1024×1024 feature map, which is then reconstructed into a four-dimensional matrix. This is the estimated value of the original CSI matrix.

[0072] Step 7: Use an end-to-end training method to train the two paradigms separately. The first paradigm jointly trains the encoder and decoder, and the second paradigm only trains the decoder. The goal is to minimize the difference between the original CSI matrix and the matrix estimate output by the decoder, and preserve the feedback network parameters trained under the two paradigms.

[0073] The first feedback paradigm uses the Xavier method for initialization. The second feedback paradigm designs an initialization method based on the corresponding sampling strategy, with the goal of padding it with zeros and initializing it to a matrix with the same dimension as the original CSI and values ​​of 0 at non-sampling positions. For the sampling strategy of this embodiment, the initialization matrix A1 is designed and expressed as

[0074]

[0075] Its dimension is 1024 × 1024. By left-multiplying A1 by the sampling matrix, we can obtain a matrix with the same dimension as the original CSI and the value of 0 at the non-sampling position, thus completing the initialization.

[0076] The loss function of the feedback network training of the two feedback paradigms uses the mean square error function, which is expressed as

[0077]

[0078] Where E(·) represents the mean, Represents the Euclidean norm. The model training optimizer uses the Adam algorithm to optimize the network based on the loss function. During the optimization process, a learning rate of 0.001 is used. Each iteration uses 200 samples from the training set to calculate the gradient, and the Adam algorithm is used to update the gradient. The entire training set is trained 1000 times. After training, the feedback network parameters under both paradigms are saved.

[0079] Step 8: Select a feedback paradigm for the two trained paradigms. Set the paradigm selection time. In the first half of the time, the user end executes the first feedback paradigm, feeds back the CSI matrix to the base station, and then measures and saves the downlink quality of this period. In the second half of the time, the second feedback paradigm is executed. After the feedback, the downlink quality of this period is measured and saved. The downlink quality of the two periods is compared, and the paradigm with better performance is determined as the subsequent CSI feedback method.

[0080] Step 9: Set the interval time and measure the downlink quality of the user end at each interval. If the quality drops significantly, reselect the feedback paradigm; otherwise, the feedback paradigm remains unchanged.

[0081] Compared with the prior art, the present invention has the following advantages:

[0082] (1) This paper extends the CSI feedback problem based on massive MIMO systems to ultra-large-scale MIMO scenarios, establishes an ultra-large-scale MIMO channel model, obtains its downlink CSI matrix, and randomly generates different CSI matrices by changing the antenna array departure angle to form a data set, thus filling the gap in the available CSI data set in ultra-large-scale MIMO scenarios.

[0083] (2) In the first feedback paradigm of the present invention, the codec shares the same network training parameters. The second feedback paradigm uses a sampling strategy to extract the complete CSI corresponding to some subcarrier frequencies for feedback. At the base station, the correlation between the subcarriers of the ultra-large-scale MIMO channel is used to predict the estimated value of the original CSI matrix. No encoder is required. Both paradigms can effectively reduce network training parameters and are simple and easy to implement.

[0084] (3) The present invention divides the complex CSI matrix into a real matrix and an imaginary matrix. Utilizing the correlation between the real and imaginary numbers in the CSI matrix of ultra-large-scale MIMO, only the real part is trained, and the imaginary part is used for verification. The real and imaginary parts share the same training network for compression and feedback, reducing the network computational scale by only one order of magnitude.

[0085] (4) The codec and sampling strategy in the present invention can be adjusted according to specific applications, and the method has strong scalability and wide adaptability.

[0086] (5) A method for generating an original CSI matrix data set is provided: The present invention extends the CSI feedback problem based on a large-scale MIMO system to an ultra-large-scale MIMO scenario, establishes an ultra-large-scale MIMO channel model, obtains its downlink CSI matrix, and randomly generates different CSI matrices by changing the antenna array departure angle to form a data set, thereby filling the gap in the available CSI data set in the ultra-large-scale MIMO scenario.

[0087] Example 2

[0088] This embodiment provides an electronic device, including: one or more processors and a memory, wherein the memory stores one or more programs, and the one or more programs include instructions for executing the method described in Embodiment 1.

[0089] Example 3

[0090] This embodiment provides a computer-readable storage medium, including one or more programs for execution by one or more processors of an electronic device, wherein the one or more programs include instructions for executing the method described in Example 1.

[0091] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and such modifications or substitutions are intended to be within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be subject to the scope of protection of the claims.

Claims

1. A channel state information feedback method for ultra-large-scale MIMO, characterized in that: Applied to the base station, it includes the following steps: Sending a pilot signal to a user terminal, receiving compressed first information based on a first feedback paradigm in a first time period, and receiving second information based on a second feedback paradigm obtained by sampling complete CSI corresponding to a subcarrier frequency in a second time period; Decompressing the first information and the second information respectively, obtaining CSI matrix estimates of the ultra-large-scale MIMO in the first time period and the second time period, and based on downlink quality in the first time period and the second time period, performing CSI matrix feedback using a feedback paradigm with better downlink quality in subsequent communication processes; The compression of the first information is achieved by an encoder that is set at the user end and pre-trained based on the original CSI matrix data set, the second information is obtained based on the sampling strategy of the user end, and the decompression is achieved by a decoder that is set at the base station end and pre-trained based on the original CSI matrix data set, and the encoder and decoder of the first feedback paradigm share the same network training parameters.

2. The channel state information feedback method for ultra-large-scale MIMO according to claim 1, characterized in that: The original CSI matrix data set is divided into a real data set and an imaginary data set, the real data set is used for training and verification, and the real data set and the imaginary data set are used for testing at the same time.

3. The channel state information feedback method for ultra-large-scale MIMO according to claim 1, characterized in that: The two feedback paradigms are trained with a fixed or variable learning rate based on the optimizer using the mean squared error function as the loss function.

4. The channel state information feedback method for ultra-large-scale MIMO according to claim 1, characterized in that: The training process of the encoder and the decoder includes the following steps: To minimize the difference between the original CSI matrix and the estimated CSI matrix output by the decoder, an end-to-end training method is used to train the two feedback paradigms respectively, and the feedback network parameters trained under the first feedback paradigm and the second feedback paradigm are saved.

5. The channel state information feedback method for ultra-large-scale MIMO according to claim 4, characterized in that: The specific training of the two feedback paradigms is as follows: The encoder and decoder are jointly trained for the first feedback paradigm, while only the decoder is trained for the second paradigm.

6. The channel state information feedback method for ultra-large-scale MIMO according to claim 1, characterized in that: The encoder and decoder are implemented based on a long short-term memory recurrent neural network.

7. The channel state information feedback method for ultra-large-scale MIMO according to claim 1, characterized in that: The acquisition of the original CSI matrix data set includes the following steps: A very large-scale MIMO channel model is constructed, and fixed parameters except the Euler angle at the center of the transmit antenna array are set. Multiple CSI matrices are generated by changing the Euler angle at the center of the transmit antenna array to form the original CSI matrix data set.

8. The channel state information feedback method for ultra-large-scale MIMO according to claim 7, characterized in that: The fixed parameters include the number of transmitting and receiving antenna subarrays, the number of antenna elements in the transmitting and receiving antenna subarrays, the three-dimensional coordinates of the centers of the transmitting and receiving antenna arrays, the Euler angle of the center of the receiving antenna array, the transmitting frequency, the bandwidth, the number of carriers, and the number of subbands in each subcarrier.

9. The channel state information feedback method for ultra-large-scale MIMO according to claim 1, characterized in that: The channel state information feedback method further includes: The downlink quality is detected at fixed intervals. If the quality drops beyond a certain threshold, a better feedback mode is selected for CSI matrix feedback.

10. The channel state information feedback method for ultra-large-scale MIMO according to claim 1, characterized in that: The first feedback paradigm and the second feedback paradigm are set in parallel, and the base station is provided with two decoders with the same structure for decompressing the first information and the second information respectively.

Images