Channel acquisition method of joint optimization of pilot compressed channel estimation and channel feedback

Abstract

The application provides a channel acquisition method for joint optimization of pilot compression channel estimation and channel feedback, and the steps are as follows: based on an N*1 large-scale MIMO antenna system, a deep neural network is used to realize pilot compression, channel estimation, channel compression quantization and feedback channel information recovery, and the modules are jointly optimized through learning training, and the trained pilot compression, channel estimation, channel compression quantization and feedback channel information recovery neural network is used to realize channel information acquisition of the sending end. Through the multi-stage learning training and joint optimization of the pilot compression, channel estimation, channel compression quantization and feedback channel information recovery modules by using the DNN, the overall optimization of the system performance is realized, and meanwhile, the problem of learning falling into local optimization caused by single-stage learning is avoided, and the optimal channel information acquisition under low pilot overhead is realized.

Description

Technical Field

[0001] This invention relates to the field of wireless mobile communication system technology, and in particular to a channel acquisition method that combines pilot compression channel estimation and channel feedback joint optimization. Background Technology

[0002] Massive MIMO (Multi-User Multiple-Operation) antenna arrays can simultaneously support more users transmitting and receiving signals, thereby increasing the channel capacity and data traffic of mobile networks by tens of times or more, while drastically reducing interference between multiple users. Therefore, it has received continuous and high attention from researchers since its inception. To achieve the above advantages in massive MIMO systems, effective precoding (beamforming) is required at the transmitting end, i.e., the rational design of the signal transmission coefficients of each antenna element in the massive MIMO. Massive MIMO precoding (beamforming) requires accurate channel state information to be obtained at the transmitting end. To obtain this channel information, in FDD systems, the base station first broadcasts pilot sequences via the downlink. Users use the received pilot sequences to estimate their channel and feed back the quantized channel information to the base station, thus completing the channel information acquisition work at the transmitting end.

[0003] However, in massive MIMO systems, with the dramatic increase in antenna size, both pilot and channel feedback overhead increase by orders of magnitude. This has become one of the key bottlenecks restricting the large-scale commercialization of massive MIMO. To reduce pilot overhead, compressed pilot transmission is crucial. Such channel information acquisition systems, such as... Figure 1 As shown, the transmitter first transmits the low-overhead pilot signal through the various antenna elements of the massive MIMO using a designed pilot compression transmission module. After the user receives the pilot signal using the receiving antenna, the channel information cannot be directly obtained because the pilot signal has been compressed. The user needs a specially designed channel estimation module to estimate the channel information of the massive MIMO. Finally, the user uses a specially designed channel feedback module to quantize the estimated channel into multiple bits and feed it back to the transmitter, which then replies with the channel information, thus completing the channel information acquisition process for the transmitter. In such a system, although the pilot transmission, channel estimation, and channel information feedback processes are highly coupled, they are designed independently and cannot achieve optimal overall system performance.

[0004] Existing large-scale and ultra-large-scale MIMO channel information acquisition systems all have inherent flaws. In these systems, although pilot transmission, channel estimation, and channel information feedback processes are highly coupled, they are designed independently. For example, in the design of pilot transmission and channel estimation, to reduce overhead, the number of pilot transmissions should be minimized while ensuring channel estimation performance. However, this process does not consider the channel feedback error, which is the final performance factor of the entire acquisition process. Therefore, the designed pilot overhead may result in channel estimation performance being far better than channel feedback performance, leading to pilot waste. Alternatively, the designed pilot overhead may result in channel estimation performance being far worse than channel feedback performance, leading to feedback bit waste. This causes the entire system design to deviate significantly from the optimal state. Summary of the Invention

[0005] To address the issue that individual design of each module in methods for obtaining channel information for large-scale and ultra-large-scale MIMO cannot achieve optimal overall system performance, this invention proposes a channel acquisition method that jointly optimizes pilot compression, channel estimation, and channel feedback. Pilot compression, channel estimation, channel compression quantization, and feedback channel information recovery are all implemented using a deep neural network (DNN), and these modules are jointly optimized through a learning and training method.

[0006] The technical solution of this invention is implemented as follows:

[0007] A channel acquisition method combining pilot compression channel estimation and channel feedback joint optimization is proposed, based on an N×1 massive MIMO antenna system. The channel from the N transmit antennas to the receive antennas is represented as h = [h0 h1 … h N-1 The pilot compression, channel estimation, channel compression quantization, and feedback channel information recovery modules are optimized through multi-stage learning using neural networks; the steps are as follows:

[0008] Step 1: Pilot compression module learning and training: The transmitter has N transmitting antennas, N... p One pilot resource; using a i,j Let a represent the pilot symbol transmitted by the i-th transmitting antenna on the j-th pilot resource, then all a i,j Combined into a dimension of N×N p The pilot transmission matrix A is given by A = FW, where W is the pilot compression matrix and F is an N-dimensional DFT matrix. W is used as a layer of neural network and is determined by learning and training with channel information.

[0009] Step 2: Learning and training of the channel estimation module: The receiving antenna is in N p The received signal on the pilot resource, the channel estimation neural network at the receiver end, is based on the receiving antenna in N pThe received signal on the pilot resources is used to recover the channel information from each transmit antenna to the receive antenna; the channel estimation neural network at the receiver is trained, and the input of the trained channel estimation neural network at the receiver is N. p The received signal on each pilot resource is output as an estimate of the channel from N transmit antennas to the receive antennas.

[0010] Step 3: Learning and training of the channel compression quantization module and the feedback channel information recovery module: Keeping the parameters of the neural networks of the pilot compression module and the channel estimation module unchanged, the estimated values ​​of the channels from the N transmit antennas to the receive antennas are used to generate channel quantization bits through the channel quantization neural network; the receiver feeds back the channel quantization bits to the transmitter, and finally the transmitter uses the feedback channel information recovery neural network to recover the original channel estimate based on the feedback bits.

[0011] Step 4: Joint Training: The parameters obtained in Step 1 are used as the initial values ​​of the neural network parameters for the pilot compression module, the parameters obtained in Step 2 are used as the initial values ​​of the neural network parameters for the channel estimation module, and the parameters obtained in Step 3 are used as the initial values ​​of the neural network parameters for the channel quantization neural network module and the feedback channel information recovery module. The four modules are jointly trained using a large amount of channel data. The trained neural networks for pilot compression, channel estimation, channel compression quantization, and feedback channel information recovery are used to achieve channel information acquisition at the transmitting end.

[0012] A channel acquisition method combining pilot compression channel estimation and channel feedback joint optimization is proposed, based on an N×1 massive MIMO antenna system. The channel from the N transmit antennas to the receive antennas is represented as h = [h0 h1 … h N-1 The pilot compression, channel compression quantization, and feedback channel information recovery modules are optimized using neural networks; the steps are as follows:

[0013] Step 1: Pilot compression module learning and training: The transmitter has N transmitting antennas, N... p One pilot resource; using b i,j Let represent the pilot symbol transmitted by the i-th transmitting antenna on the j-th pilot resource, then all b i,j Combined into a dimension of N×N p The observation matrix B, B = FW', is used, where W' is the perception matrix and F is an N-dimensional DFT matrix; W' is used as a layer of neural network and is determined by learning and training with channel information.

[0014] Step 2: Learning and training of the channel compression and quantization module: N signals transmitted from N transmit antennas. p After ×N symbols pass through the channel, the receiving antenna is at N pThe receiver receives signals on pilot resources; the receiver generates channel quantization bits for channel feedback using a channel quantization neural network and feeds them back to the transmitter.

[0015] Step 3: Learning and training of the feedback channel information recovery module: The transmitting end uses the feedback channel information recovery neural network to recover the original channel estimate based on the feedback bits;

[0016] Step 4: Use the trained neural network for pilot compression, channel compression quantization, and feedback channel information recovery to acquire channel information at the transmitting end.

[0017] The method for obtaining the channel quantization bits is as follows:

[0018] y=Sigmoid(a*x) or y=2Sigmoid(a*x)-1;

[0019] Where x is the input of the activation function, y is the output of the activation function, and a is the steepness coefficient, satisfying a>0; the activation function is used to generate channel quantization bits. If the number of feedback bits is B, then the number of outputs of the output layer of the channel quantization neural network is B.

[0020] The method for obtaining the channel quantization bits is as follows:

[0021] y=tanh(a*x) or y=(tanh(a*x)+1) / 2;

[0022] Where x is the input of the activation function, y is the output of the activation function, and a is the steepness coefficient, satisfying a>0; the activation function is used to generate channel quantization bits. If the number of feedback bits is B, then the number of outputs of the output layer of the channel quantization neural network is B.

[0023] The method for obtaining the channel quantization bits is as follows:

[0024] y=a*exp(b*x) / (1+a*exp(b*x)) or y=2*a*exp(b*x) / (1+a*exp(b*x))-1;

[0025] Where x is the input of the activation function, y is the output of the activation function, a and b are steepness coefficients, and a>0 and b>0 are satisfied; the activation function is used to generate channel quantization bits. If the number of feedback bits is B, then the number of outputs of the output layer of the channel quantization neural network is B.

[0026] The method for obtaining the channel quantization bits is as follows:

[0027] y=(a*exp(b*x)-1) / (1+a*exp(b*x)) or y=((a*exp(b*x)-1) / (1+a*exp(b*x))+1) / 2;

[0028] Where x is the input of the activation function, y is the output of the activation function, a and b are steepness coefficients, and a>0 and b>0 are satisfied; the activation function is used to generate channel quantization bits. If the number of feedback bits is B, then the number of outputs of the output layer of the channel quantization neural network is B.

[0029] The method for obtaining the channel quantization bits is as follows:

[0030] y = (a*x) b ) / (1+a*((x*Relu(x)) / 2)) b ) or y = ((a*x) b ) / (1+a*((x+Relu(x)) / 2)) b )+1) / 2;

[0031] Where x is the input of the activation function, y is the output of the activation function, a and b are steepness coefficients, and satisfy a>0, b>0, and b is an odd number; the activation function is used to generate channel quantization bits. If the number of feedback bits is B, then the number of outputs of the output layer of the channel quantization neural network is B.

[0032] The method for obtaining the channel quantization bits is a codebook-based quantization method:

[0033] y = softmax(a*x*W1);

[0034] Where x is the input of the activation function, y is the output of the activation function, W1 is the codebook, and a is the steepness coefficient, satisfying a>0; the activation function is used to generate channel quantization bits. If the number of feedback bits is B, then the number of outputs of the output layer of the channel quantization neural network is 2. B .

[0035] The optimization objective of each module's learning and training is to recover the output of the channel information recovery neural network. and the actual channel [h0 h1 … h N-1 The mean square error between them is the smallest:

[0036]

[0037] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0038] 1) By using DNN to perform multi-stage learning and joint optimization of modules such as pilot compression, channel estimation, channel compression quantization and feedback channel information recovery, the overall performance of the system is optimized. At the same time, it avoids the problem of learning falling into local optima caused by single-stage learning, and truly achieves optimal channel information acquisition with low pilot overhead.

[0039] 2) Channel compression quantization directly utilizes the received pilot channel to generate the quantization bits required for channel feedback, reducing complexity. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 Diagram of an existing large-scale MIMO channel information acquisition system.

[0042] Figure 2 This is a diagram of the first-stage learning system for pilot compression transmission and channel estimation according to the present invention.

[0043] Figure 3 This is a diagram of the second-stage learning system of the neural network for channel quantization and channel information recovery according to the present invention.

[0044] Figure 4 This is a diagram of the third stage of the overall joint optimization learning system of the present invention.

[0045] Figure 5 This is a diagram of the pilot compression and channel feedback joint optimization system of the present invention. Detailed Implementation

[0046] 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. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0047] To address the issue that individual design of each module in methods for acquiring channel information for large-scale and ultra-large-scale MIMO cannot achieve optimal overall system performance, Example 1 proposes a channel information acquisition method based on multi-level optimization of pilot compression, channel estimation, and channel feedback. Pilot compression, channel estimation, channel compression quantization, and feedback channel information recovery are each implemented using a separate distributed neural network (DNN), and these modules are jointly optimized through a multi-stage learning and training method. Specifically, the entire learning and optimization process for pilot compression and channel feedback is divided into three stages.

[0048] The first stage of learning involves training a neural network to optimize pilot compression and channel estimation, such as... Figure 2 As shown, the massive MIMO system under consideration has N antennas at the transmitting end. Since each receiving antenna at the receiving end performs channel estimation independently, only one receiving antenna needs to be considered in the design. That is, consider an N×1 massive MIMO system. The channel from the N transmitting antennas to the receiving antennas is represented as h = [h0 h1 … h N-1 If traditional pilot design methods are used, N pilot resources are needed to estimate h. To reduce pilot overhead, the transmitter only provides N pilot resources. p N pilot resources. p Much smaller than the number of antennas N, i.e., N p << N, which greatly reduces pilot overhead.

[0049] Pilot compression transmission: The transmitter has N transmitting antennas, but only N antennas are used. p One pilot resource. Because N p << N, In order for the receiver to estimate the channel of all transmitting antennas, each transmitting antenna needs to transmit a pilot signal on each pilot resource. However, the pilot information symbols transmitted by different antennas on different pilot resources are different. If the pilot symbol transmitted by the i-th transmitting antenna on the j-th pilot resource is a... i,j This means that all a i,j Combined into a dimension of N*N p The pilot transmission matrix A, A = FW, has N elements in each column, which are transmitted through N transmit antennas. Here, W is the pilot compression matrix, and F is an N-dimensional DFT matrix; W is obtained through a learning method, using W as a layer of a neural network, trained on channel information to determine W.

[0050] Channel estimation: N p ×N symbols are transmitted by N antennas, and after passing through the channel, the receiving antennas are at N... p The received signal on the pilot resources is used by the channel estimation neural network (DNN) at the receiver to estimate the channel based on the received antenna position on N. pThe received signal on the pilot resources is used to recover the channel information from each transmit antenna to the receive antenna. Therefore, the input to the channel estimation neural network (channel estimation DNN) at the receiver is N. p The received signal on the pilot resource outputs an estimate of the channel from N transmit antennas to the receive antennas.

[0051] In the first stage of learning and training, the training data comes from different channels, and the training optimization objective (cost function) is to recover the output of the neural network from the channel information. and the actual channel [h0 h1 …h N-1 The mean square error between them is the smallest.

[0052]

[0053] After the first stage of learning the neural network parameters is completed and trained, the second stage of learning is carried out based on the network.

[0054] The second stage of learning: Building upon the first stage, the output of the neural network obtained in the first stage is used to train the channel quantization neural network and the channel information recovery neural network. The principle of the second-stage channel quantization neural network and channel information recovery neural network learning system is as follows: Figure 3 As shown.

[0055] During the second stage of learning, the neural network parameters for pilot compression and channel estimation obtained in the first stage remain unchanged. The output of the pilot compression and channel estimation neural networks is first processed by a channel quantization neural network to generate quantized bits. To generate quantized bits, this invention employs six methods for the activation function of the channel quantization output layer:

[0056] Method 1:

[0057] y=Sigmoid(a*x) or y=2Sigmoid(a*x)-1;

[0058] Where x is the input of the activation function, y is the output of the activation function, and a is the steepness coefficient, satisfying a > 0; the larger the value of a, the steeper the output of the activation function. The value of a can be adjusted as a hyperparameter or learned as a parameter. This activation function generates channel quantization bits. If the number of feedback bits is B, then the number of outputs in the output layer of the channel quantization neural network is B.

[0059] Method 2:

[0060] y=tanh(a*x) or y=(tanh(a*x)+1) / 2;

[0061] Where x is the input of the activation function, y is the output of the activation function, and a is the steepness coefficient, satisfying a > 0; the larger the value of a, the steeper the output of the activation function. The value of a can be adjusted as a hyperparameter or learned as a parameter. This activation function generates channel quantization bits. If the number of feedback bits is B, then the number of outputs in the output layer of the channel quantization neural network is B.

[0062] Method 3:

[0063] y=a*exp(b*x) / (1+a*exp(b*x)) or y=2*a*exp(b*x) / (1+a*exp(b*x))-1;

[0064] Here, x is the input to the activation function, y is the output of the activation function, and a and b are steepness coefficients, satisfying a>0 and b>0; the larger the values ​​of a and b, the steeper the output of the activation function. The values ​​of a and b can be adjusted as hyperparameters or learned as parameters. This activation function generates channel quantization bits. If the number of feedback bits is B, then the number of outputs in the output layer of the channel quantization neural network is B.

[0065] Method 4:

[0066] y=(a*exp(b*x)-1) / (1+a*exp(b*x)) or y=((a*exp(b*x)-1) / (1+a*exp(b*x))+1) / 2;

[0067] Here, x is the input to the activation function, y is the output of the activation function, and a and b are steepness coefficients, satisfying a>0 and b>0; the larger the values ​​of a and b, the steeper the output of the activation function. The values ​​of a and b can be adjusted as hyperparameters or learned as parameters. This activation function generates channel quantization bits. If the number of feedback bits is B, then the number of outputs in the output layer of the channel quantization neural network is B.

[0068] Method 5:

[0069] y = (a*x) b ) / (1+a*((x+Relu(x)) / 2)) b ) or y = ((a*x) b ) / (1+a*((x+Relu(x)) / 2)) b )+1) / 2;

[0070] Here, x is the input to the activation function, y is the output of the activation function, and a and b are steepness coefficients, satisfying a>0, b>0, and b is an odd number; the larger the values ​​of a and b, the steeper the output of the activation function. The values ​​of a and b can be adjusted as hyperparameters or learned as parameters. This activation function generates channel quantization bits. If the number of feedback bits is B, then the number of outputs in the output layer of the channel quantization neural network is B.

[0071] Method Six: A codebook-based quantization method:

[0072] y = softmax(a*x*W1);

[0073] Where x is the input of the activation function, y is the output of the activation function, W1 is the codebook, and a is the steepness coefficient, satisfying a > 0; the larger the value of a, the steeper the output of the activation function. The value of a can be adjusted as a hyperparameter or learned as a parameter. Using this activation function to generate channel quantization bits, if the number of feedback bits is B, then the number of outputs in the output layer of the channel quantization neural network is 2. B .

[0074] The receiver feeds back the quantized bits to the transmitter, and finally, the transmitter uses a channel information recovery neural network to recover the original channel estimate based on the fed-back bits. The goal of the second-stage learning and training optimization is to optimize the output of the channel information recovery neural network. and the actual channel [h0 h1 … h N-1 The mean square error between them is the smallest.

[0075]

[0076] After the parameters of the learning channel feedback neural network are trained in the second stage, the third stage of learning is carried out based on the first and second stages of learning.

[0077] The third stage of learning: Based on the first and second stages of learning, the neural networks for pilot compression, channel estimation, channel quantization, and channel information recovery obtained from the first and second stages are jointly trained and optimized. The principle of the third-stage learning system is as follows: Figure 4 As shown.

[0078] During the third stage of learning, the initial parameters of the pilot compression and channel estimation neural networks are the parameters obtained in the first stage of learning, while the initial parameters of the channel quantization neural network and the channel information recovery neural network are the parameters obtained in the second stage of learning. Then, these two modules are jointly trained using a large amount of channel data.

[0079] The goal of the third stage of learning, training, and optimization is to recover the output of the channel information recovery neural network. and the actual channel [h0 h1 … hN-1 The mean square error between them is the smallest.

[0080]

[0081] After the third stage of learning and training is completed, the trained neural network for pilot compression, channel estimation, channel compression quantization, and feedback channel information recovery is used to acquire channel information at the transmitting end.

[0082] In Example 1, pilot compression, channel estimation, channel compression quantization, and feedback channel information recovery are each implemented using a separate neural network (DNN), and these modules are jointly optimized through a multi-stage learning and training method. This scheme solves the problem that individual design of each module in large-scale and ultra-large-scale MIMO channel information acquisition methods cannot achieve optimal overall system performance, thus realizing overall system performance optimization. It also avoids the problem of learning falling into local optima caused by single-stage learning, truly achieving optimal channel information acquisition with low pilot overhead.

[0083] To address the issue that individual module designs in methods for acquiring channel information for large-scale and ultra-large-scale MIMO cannot achieve optimal overall system performance, and to reduce the complexity of multi-level training, Example 2 proposes a channel information acquisition method based on the overall joint optimization of pilot compression channel estimation and channel feedback. This scheme eliminates the channel estimation module and utilizes channel compression quantization to directly generate the quantization bits required for channel feedback from the received pilot channel, thereby reducing complexity. Specifically, pilot compression, channel compression quantization, and feedback channel information recovery are all implemented using a single neural network (DNN), and these modules are jointly optimized through a single, integrated learning process.

[0084] like Figure 5 As shown, consider a massive MIMO system with N antennas at the transmitting end. Since each receiving antenna at the receiving end performs channel estimation independently, we can design the system by considering only one receiving antenna. That is, consider an N*1 massive MIMO system. The channel from the N transmitting antennas to the receiving antennas is represented by h = [h0 h1 … h N-1 The symbol is used to represent this. If traditional pilot design methods are used, N pilot resources are needed to estimate h. To reduce pilot overhead, the transmitter only provides N pilot resources. p N pilot resources. p Much smaller than the number of antennas N, i.e., N p << N, which greatly reduces pilot overhead.

[0085] Pilot compression transmission: The transmitter has N transmitting antennas, N p One pilot resource; because N p<< N, In order for the receiver to estimate the channel of all transmitting antennas, each antenna at the transmitter needs to transmit a pilot signal on each pilot resource. However, the pilot information symbols transmitted by different antennas on different pilot resources are different. Using b i,j Let represent the pilot symbol transmitted by the i-th transmitting antenna on the j-th pilot resource, then all b i,j Combined into a dimension of N*N p The observation matrix B, B = FW', is obtained by learning a neural network, where W' is the perception matrix and F is an N-dimensional DFT matrix. W' is obtained by learning a neural network layer and training it with channel information to determine W'.

[0086] Channel compression quantization: N transmitted by N antennas p After N symbols pass through the channel, the receiving antenna is at N... p The receiver receives signals on pilot resources; the receiver uses a channel quantization neural network to directly generate quantized bits for channel feedback based on the received pilot signals, and feeds these quantized bits back to the transmitter. To generate the quantized bits, the output layer activation function of the channel quantization neural network in this invention employs six methods:

[0087] Method 1:

[0088] y=Sigmoid(a*x) or y=2Sigmoid(a*x)-1;

[0089] Where x is the input of the activation function, y is the output of the activation function, and a is the steepness coefficient, satisfying a > 0; the larger the value of a, the steeper the output of the activation function. The value of a can be adjusted as a hyperparameter or learned as a parameter. This activation function generates channel quantization bits. If the number of feedback bits is B, then the number of outputs in the output layer of the channel quantization neural network is B.

[0090] Method 2:

[0091] y=tanh(a*x) or y=(tanh(a*x)+1) / 2;

[0092] Where x is the input of the activation function, y is the output of the activation function, and a is the steepness coefficient, satisfying a > 0; the larger the value of a, the steeper the output of the activation function. The value of a can be adjusted as a hyperparameter or learned as a parameter. This activation function generates channel quantization bits. If the number of feedback bits is B, then the number of outputs in the output layer of the channel quantization neural network is B.

[0093] Method 3:

[0094] y=a*exp(b*x) / (1+a*exp(b*x)) or y=2*a*exp(b*x) / (1+a*exp(b*x))-1;

[0095] Here, x is the input to the activation function, y is the output of the activation function, and a and b are steepness coefficients, satisfying a>0 and b>0; the larger the values ​​of a and b, the steeper the output of the activation function. The values ​​of a and b can be adjusted as hyperparameters or learned as parameters. This activation function generates channel quantization bits. If the number of feedback bits is B, then the number of outputs in the output layer of the channel quantization neural network is B.

[0096] Method 4:

[0097] y=(a*exp(b*x)-1) / (1+a*exp(b*x)) or y=((a*exp(b*x)-1) / (1+a*exp(b*x))+1) / 2;

[0098] Here, x is the input to the activation function, y is the output of the activation function, and a and b are steepness coefficients, satisfying a>0 and b>0; the larger the values ​​of a and b, the steeper the output of the activation function. The values ​​of a and b can be adjusted as hyperparameters or learned as parameters. This activation function generates channel quantization bits. If the number of feedback bits is B, then the number of outputs in the output layer of the channel quantization neural network is B.

[0099] Method 5:

[0100] y = (a*x) b ) / (1+a*((x+Relu(x)) / 2)) b ) or y = ((a*x) b ) / (1+a*((x+Relu(x)) / 2)) b )+1) / 2;

[0101] Here, x is the input to the activation function, y is the output of the activation function, and a and b are steepness coefficients, satisfying a>0, b>0, and b is an odd number; the larger the values ​​of a and b, the steeper the output of the activation function. The values ​​of a and b can be adjusted as hyperparameters or learned as parameters. This activation function generates channel quantization bits. If the number of feedback bits is B, then the number of outputs in the output layer of the channel quantization neural network is B.

[0102] Method Six: A codebook-based quantization method:

[0103] y = softmax(a*x*W1);

[0104] Where x is the input of the activation function, y is the output of the activation function, W1 is the codebook, and a is the steepness coefficient, satisfying a > 0; the larger the value of a, the steeper the output of the activation function. The value of a can be adjusted as a hyperparameter or learned as a parameter. Using this activation function to generate channel quantization bits, if the number of feedback bits is B, then the number of outputs in the output layer of the channel quantization neural network is 2. B .

[0105] Feedback channel information recovery: The transmitting end uses a feedback channel information recovery neural network to recover the original channel estimate based on the feedback bits.

[0106] The three neural networks mentioned above need to be trained before use. During training, the training data comes from different channels, and the goal of training and optimization is to recover the output of the neural network based on the channel information. and the actual channel [h0 h1… h N-1 The mean square error between them is the smallest. That is:

[0107]

[0108] After the parameters of the pilot compression, channel quantization, and channel information recovery neural network are trained, the trained neural network is used to complete the tasks of pilot compression, channel feedback, and channel information recovery, thereby fully realizing the acquisition of channel information at the transmitter of the large-scale MIMO system.

[0109] Example 2 utilizes channel compression quantization to directly generate the quantization bits required for channel feedback from the received pilot channel, thereby reducing complexity. This scheme solves the problem that individual module designs in large-scale and ultra-large-scale MIMO channel information acquisition methods cannot achieve optimal overall system performance, thus realizing overall system performance optimization. It also reduces complexity, truly enabling optimal channel information acquisition with low pilot overhead and low complexity.

[0110] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A channel acquisition method that combines pilot compression channel estimation and channel feedback joint optimization, characterized in that, In an N×1 large-scale antenna system, the channel from the N transmitting antennas to the receiving antennas is represented as follows: The pilot compression, channel estimation, channel compression quantization, and feedback channel information recovery modules are optimized through multi-stage learning using neural networks. The steps are as follows: Step 1: Pilot compression module learning and training: The transmitter has N transmitting antennas. One pilot resource; use Let represent the pilot symbol transmitted by the i-th transmitting antenna on the j-th pilot resource, then all Combined into a dimension The pilot transmission matrix A, Where W is the pilot compression matrix and F is the N-dimensional DFT matrix; W is used as a layer of neural network, and is determined by learning and training with channel information; Step 2: Learning and training of the channel estimation module: The receiving antenna in The received signal on the pilot resources, the channel estimation neural network at the receiver end, is based on the receiving antenna... The received signal on the pilot resources is used to recover the channel information from each transmit antenna to the receive antenna; the channel estimation neural network at the receiver is trained, and the input of the trained channel estimation neural network at the receiver is... The received signal on each pilot resource is output as an estimate of the channel from N transmit antennas to the receive antennas. ; Step 3: Learning and training of the channel compression quantization module and the feedback channel information recovery module: Keeping the parameters of the neural networks of the pilot compression module and the channel estimation module unchanged, the estimated values ​​of the channels from the N transmit antennas to the receive antennas are used to generate channel quantization bits through the channel quantization neural network; the receiver feeds back the channel quantization bits to the transmitter, and finally the transmitter uses the feedback channel information recovery neural network to recover the original channel estimate based on the feedback bits. Step 4: Joint Training: The parameters obtained in Step 1 are used as the initial values ​​of the neural network parameters for the pilot compression module, the parameters obtained in Step 2 are used as the initial values ​​of the neural network parameters for the channel estimation module, and the parameters obtained in Step 3 are used as the initial values ​​of the neural network parameters for the channel quantization neural network module and the feedback channel information recovery module. The four modules are jointly trained using a large amount of channel data. The trained neural networks for pilot compression, channel estimation, channel compression quantization, and feedback channel information recovery are used to achieve channel information acquisition at the transmitting end.

2. The channel acquisition method for joint optimization of pilot compression channel estimation and channel feedback according to claim 1, characterized in that, The method for obtaining the channel quantization bits is as follows: y=Sigmoid(a*x) or y=2Sigmoid(a*x)-1; Where x is the input of the activation function, y is the output of the activation function, and a is the steepness coefficient, satisfying a>0; the activation function is used to generate channel quantization bits. If the number of feedback bits is B, then the number of outputs of the output layer of the channel quantization neural network is B.

3. The channel acquisition method for joint optimization of pilot compression channel estimation and channel feedback according to claim 1, characterized in that, The method for obtaining the channel quantization bits is as follows: y=tanh(a*x) or y=(tanh(a*x)+1) / 2; Where x is the input of the activation function, y is the output of the activation function, and a is the steepness coefficient, satisfying a>0; the activation function is used to generate channel quantization bits. If the number of feedback bits is B, then the number of outputs of the output layer of the channel quantization neural network is B.

4. The channel acquisition method for joint optimization of pilot compression channel estimation and channel feedback according to claim 1, characterized in that, The method for obtaining the channel quantization bits is as follows: y=a*exp(b*x) / (1+ a*exp(b*x)) or y=2*a*exp(b*x) / (1+ a*exp(b*x))-1; Where x is the input of the activation function, y is the output of the activation function, a and b are steepness coefficients, and a>0 and b>0 are satisfied; the activation function is used to generate channel quantization bits. If the number of feedback bits is B, then the number of outputs of the output layer of the channel quantization neural network is B.

5. The channel acquisition method for joint optimization of pilot compression channel estimation and channel feedback according to claim 1, characterized in that, The method for obtaining the channel quantization bits is as follows: y=(a*exp(b*x)-1) / (1+ a*exp(b*x)) or y=((a*exp(b*x)-1) / (1+ a*exp(b*x)) +1) / 2; Where x is the input of the activation function, y is the output of the activation function, a and b are steepness coefficients, and a>0 and b>0 are satisfied; the activation function is used to generate channel quantization bits. If the number of feedback bits is B, then the number of outputs of the output layer of the channel quantization neural network is B.

6. The channel acquisition method for joint optimization of pilot compression channel estimation and channel feedback according to claim 1, characterized in that, The method for obtaining the channel quantization bits is as follows: y=(a*x b ) / (1+ a*((x*Relu(x)) / 2) b ) or y=((a*x b ) / (1+ a*((x+Relu(x)) / 2) b ) +1) / 2; Where x is the input of the activation function, y is the output of the activation function, a and b are steepness coefficients, and satisfy a>0, b>0, and b is an odd number; the activation function is used to generate channel quantization bits. If the number of feedback bits is B, then the number of outputs of the output layer of the channel quantization neural network is B.

7. The channel acquisition method for joint optimization of pilot compression channel estimation and channel feedback according to claim 1, characterized in that, The method for obtaining the channel quantization bits is a codebook-based quantization method: y = softmax(a*x*W1); Where x is the input of the activation function, y is the output of the activation function, W1 is the codebook, and a is the steepness coefficient, satisfying a>0; the activation function is used to generate channel quantization bits. If the number of feedback bits is B, then the number of outputs of the output layer of the channel quantization neural network is 2. B .

8. The channel acquisition method for joint optimization of pilot compression channel estimation and channel feedback according to claim 1, characterized in that, The optimization objective of each module's learning and training is to recover the output of the channel information recovery neural network. and actual channel The mean square error between them is the smallest: 。 9. A channel acquisition method that combines pilot compression channel estimation and channel feedback joint optimization, characterized in that, In an N×1 large-scale antenna system, the channel from the N transmitting antennas to the receiving antennas is represented as follows: The pilot compression, channel compression quantization, and feedback channel information recovery modules are optimized using neural networks. The steps are as follows: Step 1: Pilot compression module learning and training: The transmitter has N transmitting antennas. One pilot resource; used Let represent the pilot symbol transmitted by the i-th transmitting antenna on the j-th pilot resource, then all Combined into a dimension The observation matrix B, ,in, Let F be the perceptual matrix and F be the N-dimensional DFT matrix; As a layer of neural network, it is determined through learning and training with channel information. ; Step 2: Learning and training of the channel compression and quantization module: transmitted by N transmit antennas. After each symbol passes through the channel, the receiving antenna... The receiver receives signals on pilot resources; the receiver generates channel quantization bits for channel feedback using a channel quantization neural network and feeds them back to the transmitter. Step 3: Learning and training of the feedback channel information recovery module: The transmitting end uses the feedback channel information recovery neural network to recover the original channel estimate based on the feedback bits; Step 4: Use the trained neural network for pilot compression, channel compression quantization, and feedback channel information recovery to acquire channel information at the transmitting end.

10. The channel acquisition method for joint optimization of pilot compression channel estimation and channel feedback according to claim 9, characterized in that, The method for obtaining the channel quantization bits is as follows: y=Sigmoid(a*x) or y=2Sigmoid(a*x)-1; Where x is the input of the activation function, y is the output of the activation function, and a is the steepness coefficient, satisfying a>0; the activation function is used to generate channel quantization bits. If the number of feedback bits is B, then the number of outputs of the output layer of the channel quantization neural network is B.

11. The channel acquisition method for joint optimization of pilot compression channel estimation and channel feedback according to claim 9, characterized in that, The method for obtaining the channel quantization bits is as follows: y=tanh(a*x) or y=(tanh(a*x)+1) / 2; Where x is the input of the activation function, y is the output of the activation function, and a is the steepness coefficient, satisfying a>0; the activation function is used to generate channel quantization bits. If the number of feedback bits is B, then the number of outputs of the output layer of the channel quantization neural network is B.

12. The channel acquisition method for joint optimization of pilot compression channel estimation and channel feedback according to claim 9, characterized in that, The method for obtaining the channel quantization bits is as follows: y=a*exp(b*x) / (1+ a*exp(b*x)) or y=2*a*exp(b*x) / (1+ a*exp(b*x))-1; Where x is the input of the activation function, y is the output of the activation function, a and b are steepness coefficients, and a>0 and b>0 are satisfied; the activation function is used to generate channel quantization bits. If the number of feedback bits is B, then the number of outputs of the output layer of the channel quantization neural network is B.

13. The channel acquisition method for joint optimization of pilot compression channel estimation and channel feedback according to claim 9, characterized in that, The method for obtaining the channel quantization bits is as follows: y=(a*exp(b*x)-1) / (1+ a*exp(b*x)) or y=((a*exp(b*x)-1) / (1+ a*exp(b*x)) +1) / 2; Where x is the input of the activation function, y is the output of the activation function, a and b are steepness coefficients, and a>0 and b>0 are satisfied; the activation function is used to generate channel quantization bits. If the number of feedback bits is B, then the number of outputs of the output layer of the channel quantization neural network is B.

14. The channel acquisition method for joint optimization of pilot compression channel estimation and channel feedback according to claim 9, characterized in that, The method for obtaining the channel quantization bits is as follows: y=(a*x b ) / (1+ a*((x*Relu(x)) / 2) b ) or y=((a*x b ) / (1+ a*((x+Relu(x)) / 2) b ) +1) / 2; Where x is the input of the activation function, y is the output of the activation function, a and b are steepness coefficients, and satisfy a>0, b>0, and b is an odd number; the activation function is used to generate channel quantization bits. If the number of feedback bits is B, then the number of outputs of the output layer of the channel quantization neural network is B.

15. The channel acquisition method for joint optimization of pilot compression channel estimation and channel feedback according to claim 9, characterized in that, The method for obtaining the channel quantization bits is a codebook-based quantization method: y=softmax(a*x*W1); Where x is the input of the activation function, y is the output of the activation function, W1 is the codebook, and a is the steepness coefficient, satisfying a>0; the activation function is used to generate channel quantization bits. If the number of feedback bits is B, then the number of outputs of the output layer of the channel quantization neural network is 2. B .

16. The channel acquisition method for joint optimization of pilot compression channel estimation and channel feedback according to claim 9, characterized in that, The optimization objective of each module's learning and training is to recover the output of the channel information recovery neural network. and actual channel The mean square error between them is the smallest: 。

Images