A deep learning-based encoder-decoder decoupled CSI feedback method

Abstract

The application provides an encoder-decoder decoupling CSI feedback method based on deep learning, which comprises the following steps: 1. generating a CSI data set based on a feature vector; 2. training an encoder-decoder at a base station end; and 3. training an encoder at the base station end. The encoder-decoder decoupling CSI feedback method based on deep learning has the advantages that the CSI feedback can adapt to a large-scale MIMO system, the feedback overhead is reduced while the CSI recovery accuracy is ensured, the model algorithm of a manufacturer is protected, the model encoder and the model decoder of different manufacturers can be decoupled and used together, the model configuration of the terminal and the base station end is more flexible, and the CSI recovery performance is ensured.

Description

TECHNICAL FIELD

[0001] The present application relates to a decoder decoupling channel state information feedback method, and more particularly to a deep learning-based encoder-decoder decoupling channel state information feedback method. BACKGROUND

[0002] In order to understand the deep learning-based encoder-decoder decoupling channel state information (CSI) feedback method, it is necessary to first introduce the basic theoretical knowledge about CSI feedback technology, massive multiple-input-multiple-output (MIMO) technology and deep learning autoencoder structure.

[0003] 1. Massive MIMO technology

[0004] MIMO technology, namely multiple-input-multiple-output technology, refers to using multiple transmitting antennas and receiving antennas at the transmitting end and the receiving end respectively, so that the signal is transmitted and received through multiple antennas at the transmitting end and the receiving end, thereby improving the communication quality. This technology can fully utilize the spatial resources, and through multiple antennas, it realizes multiple transmission and multiple reception, which can multiply the system channel capacity without increasing the spectrum resources and antenna transmission power, showing obvious advantages, and is regarded as the core technology of the next generation of mobile communication. The massive MIMO technology has the following advantages: (1) improving the spectral efficiency and spatial resolution, the extremely high spatial degree of freedom can satisfy multiple users to communicate at the same time in the same time-frequency resource; (2) improving the energy efficiency, the use of large-scale antenna array improves the array gain, which can use smaller transmission power to achieve better communication quality, so that the system energy efficiency is improved by several orders of magnitude; (3) reducing the system complexity; (4) reducing the system deployment cost, the extremely high spatial degree of freedom of MIMO can reduce the peak-to-average ratio of the transmitted signal, so that low linear complexity, low cost and low power consumption devices can be used in the radio frequency front end, which greatly reduces the deployment cost.

[0005] 2. CSI feedback technology

[0006] In a MIMO wireless communication system, by precoding or beamforming on multiple transmit antennas, the transmission efficiency and reliability can be improved. In order to achieve high-performance precoding or beamforming, the precoding matrix or beamforming vector needs to be well matched to the channel, which requires the transmitter to obtain the channel state information better. Therefore, CSI feedback is a key technology for implementing high-performance precoding or beamforming in a MIMO system. At the same time, when performing CSI feedback, quantized feedback of the channel matrix will bring a large feedback overhead, especially when supporting a large bandwidth CSI feedback, therefore, the feedback overhead is an important problem that limits performance improvement. The basic CSI feedback process is that the base station sends pilots to the terminal, the terminal estimates the channel according to the pilots to obtain CSI data, the terminal compresses and quantizes the CSI data, and transmits it back to the base station through the feedback link. After the base station receives the compressed and quantized bit stream data, it performs dequantization and reconstruction to obtain the recovered CSI data for subsequent beamforming operations.

[0007] 3. Deep learning autoencoder

[0008] Deep learning belongs to a part of machine learning, and unlike machine learning, deep learning does not require human intervention when extracting data features, but relies on a neural network to automatically learn and extract data features. Among them, the autoencoder is a classic model structure in deep learning. The autoencoder belongs to unsupervised learning and includes an encoder and a decoder, which is a neural network that makes the input data and the output data as identical as possible. The encoder is used to compress the input data to achieve dimensionality reduction of the data. The decoder reconstructs the compressed data to make the recovered data as identical as possible to the original input data. From the function of the autoencoder, it is basically the same as the CSI feedback process, both of which first compress and reduce the dimension of the data, and then transmit it to the other end for reconstruction and recovery. Therefore, the combination of deep learning and CSI feedback technology has become a research hotspot in wireless communication, which improves the recovery accuracy of CSI feedback in MIMO systems through deep learning, and at the same time reduces the feedback overhead.

[0009] 4. Current system deficiencies and shortcomings

[0010] For a large-scale MIMO system under frequency division duplexing (FDD), the uplink and downlink work on different frequencies, so the downlink CSI is obtained by the user end and transmitted back to the base station through the feedback link. Both the traditional CSI feedback method based on codebook and the one based on compressed sensing will increase the computational complexity with the increase of the number of antennas in the system, affecting the real-time performance of the actual communication system. Therefore, the traditional CSI feedback scheme cannot be well applied to large-scale MIMO systems.

[0011] Therefore, it is considered to combine deep learning with CSI feedback, which can effectively reduce the overhead brought by feedback while improving the accuracy of CSI feedback recovery. At present, the CSI feedback based on deep learning mainly applies the structure of autoencoder, and the encoder is configured at the terminal to compress data; and the decoder is configured at the base station end to receive the compressed data sent by the terminal and reconstruct. However, in actual application, the problem of different manufacturers of terminals and base stations may be encountered. In order to protect the technology, the model based on deep learning designed by the manufacturer may not be open, so that the terminal of other manufacturers cannot configure the corresponding encoder in time. SUMMARY

[0012] In view of the defects in the prior art, the application provides an encoder-decoder decoupling CSI feedback method based on deep learning,

[0013] The CSI feedback method comprises the following steps:

[0014] Step 1. Generation of CSI data set based on feature vector:

[0015] Step 1.1. Sub-band division of full channel information data. The pilot is sent from the base station end to the terminal, and the terminal performs ideal channel estimation according to the pilot to obtain full channel information data H. Channel estimation is obtained by comparing the reference signal sent by the base station end with the signal received by the terminal, and ideal channel estimation is the completely known channel information which can be directly used as channel data. Then, the obtained full channel information data H is divided into K sub-bands, and each sub-band contains N sub-carriers, wherein the channel data of the kth sub-band is denoted as H sc . k ;

[0016] Step 1.2. Calculate the correlation matrix of the channel matrix in each sub-band; and accumulate and sum the correlation matrix to obtain the average, which is expressed as formula (1):

[0017]

[0018] In the above formula (1), H k,i is the channel matrix of the ith sub-carrier in the kth sub-band, R k is the correlation matrix of the kth sub-band, and the correlation matrix of the K sub-bands is as follows:

[0019] R=[R1,R2,…,R K ]......(2),

[0020] Step 1.3 Eigenvalue decomposition is performed on the correlation matrix to obtain the eigenvectors of each subband, and the calculation expression of the eigenvectors of each subband is as follows formula (3):

[0021] R k w k = λ k w k .......(3),

[0022] In the above formula (3): λ k is the maximum eigenvalue of the kth subband, w k is the eigenvector of the kth subband, and the CSI data based on the eigenvector is obtained, which is expressed as shown in the following formula (4):

[0023]

[0024] Step 2. Train the encoder and decoder at the base station end:

[0025] Step 2.1 The base station end trains the CSI feedback model according to the generated eigenvector-based CSI data, including the encoder f E (·) and the decoder f D (·), during the training process, the encoder first compresses the eigenvector-based CSI data w as input to obtain compressed data x, x is first uniformly quantized, and the uniform quantization is bit stream data s, and then the bit stream data s is transmitted to the decoder end, after the decoder end receives the bit stream data s, the decoder decodes and reconstructs the bit stream data s to obtain the recovered data denoted as the following formula (5):

[0026]

[0027] In the above formula (5): the parameter set θ E and θ D of the encoder and the decoder are updated with the training process, as shown in the following formula (6):

[0028]

[0029] In the above formula (6): L(·,·) is the loss function of the updated parameter set θ E and θ D , and the cosine similarity is used to evaluate the performance of the recovered accuracy of the CSI feedback, and the square of the cosine similarity is used as the loss function (Square of generalized cosine similarity, SGCS) when designing the loss function of the model training, which is expressed as the following formula (7),

[0030]

[0031] Step 2.2. After the training is completed, the CSI data w based on the feature vector is compressed by the encoder of the model to obtain the compressed quantized data x, and w and x are transmitted to the terminal for the next training, and the trained decoder f D (·) is configured at the base station end;

[0032] Step 3. The base station terminal trains the encoder:

[0033] Step 3.1. The base station terminal designs the encoder h E (·) according to the received CSI data w based on the feature vector and the compressed quantized data x; E (·) In view of the limited computing power and power consumption of the terminal, the model is simple when designing h E (·), and the MixerNet encoder or EVCsiNet encoder is considered as the terminal encoding;

[0034] Step 3.2. The encoder h E (·) is trained by inputting the CSI data w based on the feature vector as the original data and the compressed quantized data x as the label into the encoder h E (·); The process is denoted by the following formula (8):

[0035]

[0036] In the above formula (8), the parameter set θ E ′ of the encoder is continuously updated with the training process and is expressed by the following formula (9):

[0037]

[0038] In the above formula (9), L′(·) is the loss function of the updated parameter set θ E ′, and the loss function of the encoder h E (·) adopts the mean squared error loss function (MSE), and the expression is as follows:

[0039]

[0040] In the above formula (10), N is the number of data samples, x i represents the label corresponding to the i-th input data sample of the encoder h E (·), represents the output corresponding to the i-th input data sample of the encoder h E (·), E (·) is configured at the terminal.

[0041] The CSI feedback method has the following superior technical effects:

[0042] 1. The CSI feedback method uses a deep learning-based method, which enables the CSI feedback to adapt to large-scale MIMO systems, reduces the feedback overhead while ensuring the accuracy of CSI recovery.

[0043] 2. The encoder-decoder decoupled CSI feedback method proposed by the CSI feedback method considers the protection of the model algorithm by manufacturers, so that the model encoder and decoder of different manufacturers can be decoupled for use, ensuring the CSI recovery performance while making the model configuration of the terminal and the base station more flexible.

[0044] 3. Compared with the previous deep learning-based CSI feedback model, the terminal and the base station can only be equipped with fixed model encoders and decoders, while the encoder-decoder decoupled CSI feedback method proposed by the CSI feedback method makes the model configuration more flexible on both ends, while taking into account the protection of the model algorithm by manufacturers, and only data is transmitted between the terminal and the base station, not the model, which reduces the overhead caused by transmission. BRIEF DESCRIPTION OF DRAWINGS

[0045] Figure 1 is a flowchart of the CSI feedback method. DETAILED DESCRIPTION

[0046] In order to more clearly understand the above-mentioned purposes, features and advantages of the present application, the specific embodiments of the CSI feedback method of the present application will be further described in detail below in combination with the drawings and specific embodiments.

[0047] As shown in Figure 1 , the CSI feedback method configures the terminal-trained encoder and the base station-trained decoder on both sides, respectively, to realize the encoder-decoder decoupled CSI feedback method. In the actual communication process, the base station sends pilots to the terminal, and the terminal obtains the full channel information data H through ideal channel estimation, and then generates the corresponding CSI data w based on the eigenvector, which is used as the input of the terminal encoder h E (·) to obtain compressed quantization data to obtain bit stream data through uniform quantization, and transmit it to the base station through the feedback link. The base station takes the bit stream data as the decoder f D ​input of (·), dequantizes and reconstructs it to obtain recovered feature vector-based CSI data for subsequent beamforming.

[0048] The present application is not limited to the above-described embodiments, and any modification, improvement, replacement that can be conceived by those skilled in the art without departing from the essential content of the present application falls within the scope of protection of the present application.

Claims

1. A deep learning based encoder-decoder decoupled CSI feedback method, characterized in that, Comprising the following steps: Step 1. Generating a CSI dataset based on eigenvectors: Step 1.

1. Subband division of full channel information data, same as full channel information data, pilot sent from base station to terminal, terminal performs ideal channel estimation according to pilot to obtain full channel information data Channel estimation is based on comparison of reference signal sent from base station and signal received by terminal to obtain channel information, while ideal channel estimation is full known channel information which can be directly used as channel data; next, subband division is performed on obtained full channel information data into K subbands and each subband contains subcarriers, wherein channel data of the k th subband is denoted as ; Step 1.2 Calculate the correlation matrix of the channel matrix in each subband; and accumulate and average the correlation matrix, the expression is as follows (1): ......(1)， In the above formula (1): For the first k The first in the individual belt i Channel matrix of subcarriers, For the first k The related matrix of each sub-band, then K The correlation matrix of each sub-band is as follows (2): ......(2)， Step 1.3 Perform eigenvalue decomposition on the correlation matrix to obtain the eigenvector of each subband, and the calculation expression of the eigenvector of each subband is as follows (3): .......(3)， In the above equation (3), is the maximum eigenvalue of the k th subband, is the eigenvalue of the k th subband, and the CSI data based on the eigenvector is obtained as shown in the following equation (4): .......(4)； Step 2. Training the encoder-decoder at the base station end: Step 2.1 The base station trains a CSI feedback model based on the generated feature vector-based CSI data, including the encoder. With decoder During training, the encoder first processes CSI data based on feature vectors. The data is compressed as input to obtain the compressed data. ,right Advanced uniform row quantization, uniformly quantizing into bitstream data Then the bitstream data The data is transmitted to the decoder, where it receives the bitstream data. Then, the decoder processes the bitstream data. Dequantization and reconstruction are performed to obtain the recovered data. , denoted as the following formula (5): ......(5)， In the above equation (5): Parameter set of the encoder and the decoder and As the training process is constantly updated, as shown in the following equation (6): ......(6)， In the above equation (6): is a parameter set and The updated loss function, cosine similarity is used to evaluate the performance of the recovery accuracy of CSI feedback, and the square of the generalized cosine similarity (SGCS) is used as the loss function when designing the loss function of the model training, which is expressed as the following equation (7): ......(7)， Step 2.

2. After training is complete, the CSI data based on feature vectors will be... The compressed data to be quantized is obtained through the encoder of the model. ,Will and The trained decoder is then sent to the terminal for further training. Configured on the base station side; Step 3. Training the encoder at the base station end: Step 3.1 Base station terminal encodes the received CSI data based on the feature vector and compresses the data to be quantized to design the encoder In view of the limitation of terminal computing power and power consumption, the model is simple when designing , so MixerNet encoder or EVCsiNet encoder is considered as the terminal encoding Step 3.2 Encoder CSI data based on the feature vector As the original data, compress the data to be quantized As a label, input to the encoder Training in the encoder The output is the compressed data to be quantized, denoted as The training process is denoted as formula (8) below: ......(8)， In the above equation (8): parameter set of the encoder As the training process is constantly updated, expressed as the following equation (9): ......(9), In the above equation (9): is a parameter set The loss function of the updated encoder The loss function of the encoder adopts a mean squared error loss function (MSE), and the expression is as follows equation (10): ......(10)， In the above formula (10): N is the number of data samples, represents the label corresponding to the data sample input into the encoder i of the i-th input into the encoder represents the output corresponding to the data sample input into the encoder i of the i-th after the training is completed, the encoder is configured in the terminal.

Images