TW201944745A - Feedback method for use as a channel information based on deep learning - Google Patents

Abstract

A feedback method for use as a channel information based on deep learning includes generating a coded word by separating a channel state information into a real part matrix and an imaginary part matrix using a coding training module at a receiving end; sending the coded word to a transmission end by the receiving end; acquiring the coded word by the transmission end; transforming the coded word back into the real part matrix and the imaginary part matrix using a decoding training module by the transmission end; and transforming the real part matrix and the imaginary part matrix back into the channel state information using the decoding training module. The magnitude of the channel state information is NcX Nt. NC represents the quantity of the carriers of the OFDM system adopted and is larger than or equal to 1. Nt is the quantity of the transmission antenna at the transmission end.

Description

   Feedback method based on deep learning as channel state information   

The invention relates to a feedback method for channel state information, and in particular to a method for reducing channel time complexity based on deep learning as a feedback method for channel state information.

In recent years, with the rapid development of personal communication needs and the rapid increase of multimedia information exchange, the relatively available spectrum is very limited, making spectrum a precious resource. Therefore, Multiple-Input Multiple-Output, (MIMO) technology has received much attention. As one of the key technologies in the wireless communication field, it has beamforming capability, diversity gain capability, and multiplexing gain capability. It can be used at the transmitting and receiving ends. Using multiple antennas and related communication signal processing technology at the same time, it can provide space freedom without increasing the bandwidth, and effectively improve the system capacity and spectrum efficiency of the wireless system.

Pressing, the multiple-input multiple-output technology can roughly use two duplex modes: time-division duplex (TDD) or frequency-division duplex (FDD). Among them, the duplex of wireless communication (Duplex) technology refers to a method in which the transmitting end and the receiving end use channel access to implement two-way communication so that two communication devices can transmit data to each other. In addition, the large-scale multiple-input multiple-output (Massive MIMO) technology extended by the multiple-input multiple-output technology can greatly increase the system capacity and spectrum efficiency to support a larger number of users, so it is generally considered to be The main technology of the fifth generation wireless communication system, in which the channel reciprocity of the time division duplex technology depends on the complex calibration process, and the existing system uses the frequency division duplex technology extensively, for example, most of the existing mobile phone systems are It is the use of frequency division duplex technology, so that large-scale frequency division duplex multiple input multiple output (FDD Massive MIMO) system is an important development direction of multiple input multiple output technology.

However, for the existing large-scale frequency division duplex multiple-input multiple-output system, when the user equipment (User Equipment, UE) of a receiving end needs to feedback a channel state information (CSI) ) For a base station (BS) to which a transmitting end belongs, the channel state information is simplified to make the channel structure appear sparse, and then Compressive Sensing (CS) is used. The signal of the channel state information is compressed, and the channel state information can be restored again after the transmitting end is repeatedly iterated. Because the conventional method of using compressed sensing requires multiple iterations to restore the channel state information, the system time complexity is increased and the system performance is reduced.

In view of this, the conventional feedback method of channel state information does still need to be improved.

In order to solve the above problems, an object of the present invention is to provide a feedback method based on deep learning as channel state information, which can reduce system time complexity using deep learning to reconstruct the channel state information.

The feedback method based on deep learning as the channel state information of the present invention includes the following steps: a channel training information is divided into a real part matrix and an imaginary part matrix at a receiving end by a coding training model to generate a code word , The matrix size of the channel status information is xNt, Expressed as the number of original subcarriers of the OFDM system used, and ≧ 1, Nt indicates the number of transmitting antennas of the base station at the transmitting end, and Nt ≧ 1; the receiving end feeds back the encoded word to a transmitting end; the transmitting end obtains the encoded word, and encodes the encoding with a decoding training model The words are transformed back to the real part matrix and the imaginary part matrix; and the decoding training model is used to transform the real part matrix and the imaginary part matrix of the transmitting end back to the channel state information.

According to this, the method based on deep learning of the present invention as a channel state information feedback method can greatly compress the channel state information when transmitting the channel state information, and reconstruct the channel state information with extremely low complexity. In order to improve the acquisition efficiency of the channel state information, the system can be applied to large-scale multiple input multiple output technology to give full play to its advantages.

The convolutional neural network of the coding training model includes at least a first layer and a second layer. The first layer has two channels, and the real part matrix and the imaginary part matrix provided at the receiving end are in matrix size. Perform a convolution operation on the kernel of KxK, and perform a linear rectification function to generate a two-characteristic map. The second layer obtains a feedback parameter with a matrix size of Nx1 from the two feature map, and converts the feedback parameter into a M-dimensional vector. Code word, a data compression ratio of the channel status information Where K ≧ 3, N = NcxNtx2, where N is the product of the length, width, and number of the feature map, and Nc is the number of subcarriers obtained by the OFDM system, and ≧ Nc ≧ 1. In this way, compared with the conventional compressive sensing method, which cannot fully utilize the feature values possessed by the channel state information, the present invention can obtain the feature values of the code word from the channel state information, and has the identifiability of improving the code word. Sexual effect.

The decompression training model includes at least one first layer and two refined network layers. The first layer of the decoding training model converts the codeword back to the feedback parameter to obtain an initial estimate of the matrix size NcxNt, respectively. Each of the refined network layers includes at least one input layer and three convolution layers, and the input layer of the first refined network layer of the two refined network layers is used to input the initial layer The estimated real part matrix and the initial estimated imaginary part matrix, the first convolution layer of the first refined network layer, and the preliminary estimated real part matrix and the initially estimated imaginary part matrix are rolled with a kernel of matrix size KxK Product, and perform a linear rectification function to generate 8 feature maps, the second convolution layer performs a convolution operation on the 8 feature maps with a kernel of matrix size KxK, and performs a linear rectification function to generate 16 feature maps, The third convolution layer performs a convolution operation on the 16 feature maps with a kernel size of KxK, and performs a linear rectification function to generate 2 feature maps. The 2 feature maps generated by the third convolution layer are zeroed. Padding to make the 2 The matrix size of the feature map is the same as the matrix size of the real-estimation matrix and the imaginary-estimation matrix input to the input layer, and the matrices of the two feature maps are respectively equal to the real-estimation matrix and the initial estimation. The imaginary part matrix is added as the input of the input layer of the second refined network layer among the two refined network layers, and then the convolution operation and execution are performed sequentially through the three convolution layers of the second refined network layer. The linear rectification function converts the initial estimated real part matrix and the initial estimated imaginary part matrix back to the real part matrix and the imaginary part matrix. In this way, compared with the known compressed sensing method, the present invention can restore the channel state information through deep learning without multiple iterations, which has the effect of reducing time complexity.

Wherein, the receiver uses the coding training model to calculate the channel state information by two-dimensional discrete Fourier transform, so that the channel state information is converted from spatial frequency to angle and time based. The coding training model retains the channel state information. The value of the first Nc column is not zero to generate a truncated matrix, and the truncated matrix is split into the real part matrix and the imaginary part matrix to generate the code word. In this way, the amount of information that the receiving end needs to feedback can be reduced, which has the effect of reducing the channel state information feedback overhead.

The data compression ratio is 1/4, 1/16, 1/32, or 1/64. In this way, compared with the conventional compressed sensing method, the present invention can greatly compress the channel state information through deep learning, which has the effect of further reducing the channel state information feedback overhead.

The data compression ratio is 1/4 or 1/32. In this way, compared with the conventional compressive sensing method, the present invention can compress the channel state information when the channel state information is not a sparse matrix, can be applied to different antenna structures, and has the effect of improving the use range.

     〔this invention〕     

S1‧‧‧ encoding steps

S2‧‧‧ decoding steps

S3‧‧‧Sparse Step

L11‧‧‧First floor

L12‧‧‧Second floor

L21‧‧‧First floor

L22‧‧‧ Refined Network Layer

L221‧‧‧input layer

L222‧‧‧ Convolutional Layer

L223‧‧‧ Convolutional Layer

L224‧‧‧ Convolutional Layer

FIG. 1 is a flowchart of a method according to an embodiment of the present invention.

FIG. 2 is a schematic diagram of a system architecture for implementing an embodiment of the present invention.

In order to make the above and other objects, features, and advantages of the present invention more comprehensible, the following describes the preferred embodiments of the present invention and the accompanying drawings in detail, as follows: Please refer to FIG. The method flow chart of a preferred embodiment of the method based on deep learning as a feedback method of channel state information includes: an encoding step S1 and a decoding step S2.

The encoding step S1 is performed by a receiving end to transmit channel state information. Simplified into a coded word, the receiving end feeds back the coded word to a transmitting end. In this embodiment, the channel state information The channel state information of the downlink is transmitted from the base station of the transmitting end to the user equipment of the receiving end. Specifically, the receiver can train the channel state information with a coding training model. Split into a real part matrix and an imaginary part matrix to generate a coded word, and the receiving end feeds back the coded word to the transmitting end. Among them, the channel status information Is based on the spatial frequency, and the matrix size can be xNt, Indicates the number of original subcarriers of the OFDM system used in the present invention, and ≧ 1, Nt is the number of transmitting antennas of the base station, and Nt ≧ 1.

Among them, the coding training model uses deep learning convolutional neural networks (Convolutional Neural Networks, CNNs) for training and learning. For example, the convolutional neural network system may include at least a first layer L11 and a second layer L12. The first layer L11 and the second layer L12 may be a convolutional layer and a pool, respectively. A pooling layer or a fully-connected layer. In this embodiment, the first layer L11 may be a convolution layer, and the second layer L12 may be a fully-connected layer. The first layer L11 has two channels, and the two channels respectively provided at the receiving end of the real part matrix and the imaginary part matrix are convolved by a kernel (Kernel) with a matrix size of KxK, and perform a linear rectification function (Rectified Linear Unit , ReLU) to generate a two feature map (Feature Map), where K ≧ 3. The second layer L12 obtains a feedback parameter with a matrix size of Nx1 from the two feature maps, and converts the feedback parameter into an M-dimensional vector code word. The channel status information Data compression ratio Where N is the product of the length, width, and number of feature maps, and N = NcxNtx2, and Nc is the number of subcarriers obtained by the OFDM system, and ≧ Nc ≧ 1; M is the dimension of the codeword, and M N.

The decoding step S2 is capable of obtaining the codeword by the transmitting end and converting the codeword back to the channel state information. . Specifically, the transmitting end can convert the encoded word back to the real part matrix and the imaginary part matrix by using a decoding training model, and the decoding training model then converts the real part matrix and the imaginary part matrix back to the channel state information. To complete the decoding step S2. In this embodiment, the decoding training model system may include at least a first layer L21 and two RefineNet layers L22. The first layer L21 of the decoding training model may convert the codeword back to the feedback parameter. To obtain an initial estimated real part matrix and an initial estimated imaginary part matrix with matrix sizes NcxNt, respectively. The second layer L22 of the decoding training model then transforms the initial estimated real part matrix and the initial estimated imaginary part matrix back to the real part matrix and the imaginary part matrix. In this embodiment, the first layer L21 may be A fully connected layer.

Specifically, each of the refining network layers L22 may include at least one input layer L221 and three convolution layers L222 to L224. The input layer L221 of the first refined network layer of the two refined network layers L22 is used to input the matrix of the real part of the initial estimation and the matrix of the imaginary part of the initial estimation. Subsequently, the first convolution layer L222 of the first refined network layer L22 performs a convolution operation on the initial estimated real part matrix and the initial estimated imaginary part matrix using a kernel with a matrix size of 3x3, and performs a linear rectification function To generate 8 feature maps, the second convolution layer L223 performs a convolution operation on the 8 feature maps with a kernel size of KxK, and performs a linear rectification function to generate 16 feature maps. The third convolution layer L224 pairs The 16 feature maps are convolved with kernels of matrix size KxK, and a linear rectification function is performed to generate 2 feature maps. Zero feature padding is performed on the 2 feature maps generated by the third convolution layer L224. , So that the matrix size of the two feature maps is the same as the matrix size of the initial estimated real part matrix and the initial estimated imaginary part matrix input to the input layer L221. Subsequently, the sum of the two feature maps and the initial estimated real part matrix and the initial estimated imaginary part matrix are respectively used as the input of the input layer L221 of the second refined network layer of the two refined network layers L22. And then sequentially perform convolution operations and perform linear rectification functions through the three convolution layers L222 ~ L224 of the second refined network layer, so that the initial estimated real part matrix and the initial estimated imaginary part matrix are converted back to the real part matrix. And the imaginary part matrix.

Preferably, the present invention may further include a thinning step S3. In the sparse step S3, the receiving end uses the coding training model to train the channel state information. Calculate with two-dimensional discrete Fourier transform to make channel status information Conversion of space frequency to angle and time as the basis, so that the channel state information According to the orthogonal frequency division multiplexing multiple input multiple output (MIMO-OFDM) and the spatial frequency correlation of the channel, it can exhibit sparse characteristics in the angular delay domain (Angular-Delay Domain).

In detail, in the dimension of time delay, since the time delay between multipath arrivals is within a certain period of time, the channel status information may not be needed All time points in. The encoding training model retains the channel state information The value in the middle Nc column is not zero to generate a truncated matrix H, and the truncated matrix H is divided into the real part matrix and the imaginary part matrix to generate the code word. In this way, the amount of information that the receiving end needs to feedback can be reduced, which has the effect of reducing the channel state information feedback overhead.

For example, the present invention is based on deep learning as a channel state information feedback method, which can be applied to a large-scale frequency division duplex multiple-input multiple-output (FDD Massive MIMO) system based on the use of orthogonal frequency division multiplexing. Technical system. First, in order to train the encoding training model of the encoding step S1 and the decoding training model of the decoding step S2, the present invention can pass the COST 2100 Channel Model (see L. Liu et al., "The COST 2100 Channel Model "IEEE Wireless Commun., Vol. 19, no. 6, pp. 92-99, Dec. 2012.) to establish two channels, one of which is an indoor environment in the 5.3 GHz frequency band, and the other of the two channels One channel is an outdoor environment in the 300MHz frequency band. The two channels each use 100,000 sample data, 30,000 sample data, and 20,000 sample data as the training set, verification set, and test set of the encoding training model and the decoding training model. One transmitting end is provided with 32 transmitting antennas, and the OFDM system divides the frequency band into 1024 subcarriers.

In this embodiment, the channel state information is obtained through the thinning step S3. After conversion to angle and time, the channel state information is retained In the first 32 rows whose values are not zero, a two-feature map with a matrix size of 32x32 is generated. The second layer L12 of the encoding training model generates feedback parameters with a matrix size of 2048x1 according to the two feature maps, and converts the feedback parameters into encoding words of an M-dimensional vector to complete the training and learning of the encoding training model. In addition, the decoding training model is to transform the coded word into the feedback parameter to obtain an initial estimated real part matrix and an initial estimated imaginary matrix with a matrix size of 32x32, the initial estimated real part matrix and the initial estimation. The imaginary part matrix is further subjected to a convolution operation through the two refined network layers L22, so that the initially estimated real part matrix and the initially estimated imaginary part matrix are converted back to the real part matrix and the imaginary part matrix. The decoding training model converts the real part matrix and the imaginary part matrix back to the channel state information. .

The present invention is based on deep learning as a feedback method of channel state information, and is compared with the conventional methods such as LASSO, TVAL3, BM3D-AMP, and CsiRecovNet. In the present invention, the channel matrix is Compared with the reconstructed channel matrix, the normalized mean square error (NMSE) is used to measure the difference between the channel state information before encoding and after decoding, and the feedback channel state information is compared using the cosine similarity ρ (Consine Similarity). The comparison result can be as follows: As shown in Table 1. The formulas of the similarity between the normalized mean square error and the cosine can be shown as follows: Where H is the original truncation matrix, Is the truncated matrix after reconstruction.

among them, Represents the channel vector on the nth subcarrier; Is the reconstructed channel vector on the nth subcarrier.

The above Table 1 records the present invention and the conventional method. The corresponding normalized mean square error and cosine similarity when the data compression ratios are 1/4, 1/16, 1/32, and 1/64, and the best results are shown in bold. For example, when the compression ratio is 1/4, the channel state information is converted into a 512-dimensional vector codeword. It can be known from Table 1 that the present invention obtains the lowest normalized mean square error and the highest cosine similarity, so the present invention is significantly superior to the conventional method. Furthermore, the average running time of the present invention is 0.0035s, which is also better than the conventional LASSO 0.1828s, TVAL3 0.5717s and BM3D-AMP 0.3155s. In addition, when the base station is in an outdoor environment and the transmitting station has 16 and 48 transmitting antennas, the comparison results can be shown in Table 2. It can be seen that the present invention is also superior to the conventional method.

On the other hand, as shown in Table 3, Perform the thinning step S3 to make the channel state information Converting from spatial frequency to angle and time as the basis, without performing the sparse step S3, when the channel state information When the data compression ratios are 1/4 and 1/32, respectively, whether in an indoor environment or an outdoor environment, the performance of the present invention is better than performing the thinning step S3.

Although the present invention has been disclosed using the above-mentioned preferred embodiments, it is not intended to limit the present invention. Anyone skilled in the art can make various changes and modifications to the above embodiments without departing from the spirit and scope of the present invention. The technical scope protected by the invention, so the scope of protection of the present invention shall be determined by the scope of the appended patent application.

Claims (6)

A feedback method based on deep state learning as channel state information includes the following steps: a channel training state is divided into a real part matrix and an imaginary part matrix by a coding training model at a receiving end, to generate a code word, the The matrix size of the channel status information is xNt, Expressed as the number of original subcarriers of the OFDM system used, and ≧ 1, Nt represents the number of transmitting antennas of the base station at the transmitting end, and Nt ≧ 1; and the receiving end feeds the codeword to a transmitting end; the transmitting end obtains the codeword, and uses a decoding training model to The codeword is transformed back to the real part matrix and the imaginary part matrix; and the decoding training model is used to transform the real part matrix and the imaginary part matrix of the transmitting end back to the channel state information. The feedback method based on deep learning as channel state information according to item 1 of the scope of the patent application, wherein the convolutional neural network of the coding training model includes at least a first layer and a second layer, and the first layer has Two channels, and the real part matrix and the imaginary part matrix respectively provided at the receiving end are convolved with a kernel of matrix size KxK, and a linear rectification function is performed to generate a two-characteristic map, and the second layer is composed of the two-characteristic map To obtain a feedback parameter with a matrix size of Nx1, and convert the feedback parameter into an M-dimensional vector code word, a data compression ratio of the channel state information , Where K ≧ 3, N = NcxNtx2, where N is the product of the length, width, and number of feature maps, and Nc is the number of subcarriers obtained by the OFDM system, and ≧ Nc ≧ 1.     The feedback method based on deep learning as channel state information as described in item 2 of the scope of the patent application, wherein the decompression training model includes at least one first layer and two refined network layers, and the first layer of the decoding training model will The codeword is converted back to the feedback parameter to obtain an initial estimated real part matrix and an initial estimated imaginary part matrix with a matrix size of NcxNt, respectively. Each of the refined network layers includes at least one input layer and three convolutional layers. The two refinements The first refinement network layer input layer in the network layer is used to input the initial estimate real part matrix and the initial estimate imaginary part matrix, and the first convolution layer of the first refined network layer is adapted to the initial estimate. The real part matrix and the initial imaginary part matrix are convolved with a kernel with a matrix size of KxK, and a linear rectification function is performed to generate 8 feature maps. The second convolution layer uses the matrix size for the 8 feature maps. Perform a convolution operation on the kernel of KxK, and perform a linear rectification function to generate 16 feature maps. The third convolution layer performs a convolution operation on the 16 feature maps with a kernel of size KxK, and performs a linear rectification function to produce 2 feature maps, perform zero padding on the 2 feature maps generated by the third convolution layer, and make the matrix size of the 2 feature maps and the initial estimated real part matrix and the initial estimated imaginary matrix input to the input layer And the matrix of the two feature maps are respectively added to the matrix of the real part and the matrix of the imaginary part as the input layer of the second refined network layer of the two refined network layers. Input, and then sequentially perform convolution operation and linear rectification function through the three convolution layers of the second refined network layer, so that the initial estimated real part matrix and the initial estimated imaginary part matrix are converted back to the real part matrix. And the imaginary part matrix.         The feedback method based on deep learning as the channel state information described in item 3 of the scope of the patent application, wherein the receiving end uses the coding training model to calculate the channel state information by a two-dimensional discrete Fourier transform, so that the channel state information is obtained from The space frequency is converted into an angle and a time base. The encoding training model retains the non-zero values of the first Nc column in the channel state information to generate a truncated matrix, and the truncated matrix is split into the real part matrix and the The imaginary part matrix to generate the codeword.         The feedback method based on deep learning as channel state information described in item 4 of the scope of patent application, wherein the data compression ratio is 1/4, 1/16, 1/32, or 1/64.         The feedback method based on deep learning as channel state information as described in item 3 of the scope of patent application, wherein the data compression ratio is 1/4 or 1/32.